Building Production-Ready AI Agents on Google Cloud Run
● Day Zero: Technical
Workshop
📍 Theater 3
Prototyping an AI agent is faster than ever, but transitioning that prototype into a secure, scalable, and observable production application is where the real engineering begins.Prototyping an AI agent is faster than ever, but transitioning that prototype into a secure, scalable, and observable production application is where the real engineering begins. In this 90-minute interactive workshop, you will move beyond "Day 1" demos to master enterprise-grade AI deployment on Google Cloud’s serverless platform.
We have prepared hands-on labs so you can experience firsthand how to build and scale applications on Cloud Run - whether it’s an intelligent AI agent or a high-performance AI inference app with serverless GPUs. You can choose a lab of your interest or do them all:
- AI Assistant Agent: Build and orchestrate intelligent workflows using Google's Agent Development Kit (ADK) on Cloud Run with Firestore.
- Personal Agent with Secure Sandboxes: Safely execute untrusted, agent-generated code in highly secure Cloud Run Sandboxes that start in sub-second.
- High-Performance Serverless Inference: Run low-latency open model inference using Gemma 4 on Cloud Run GPUs.read more
Lisa ShenProduct Manager, Google
☆
Tue 9/29 · 9:00 AM - 10:30 AM
From AI Experiments to Enterprise Strategy: Build Your AI Investment Portfolio
● Day Zero: Leadership
Workshop
📍 Theater 2
Move beyond AI experimentation and make smarter decisions about where to invest. In this hands-on workshop, leaders will evaluate and prioritize AI opportunities within their own organizations, explore build vs. buy vs.Move beyond AI experimentation and make smarter decisions about where to invest. In this hands-on workshop, leaders will evaluate and prioritize AI opportunities within their own organizations, explore build vs. buy vs. partner decisions, and identify where to invest, where to wait, and what to stop. Through practical exercises and discussion, attendees will leave with a framework for building a focused AI investment strategy and a clear path toward action.read more
Evangelos SimoudisCo-founder and Managing Dir, Synapse Partners
Designing Reliable Agentic Systems that make it into Production
● Day Zero: Technical
Workshop
📍 Theater 3
Most agents don't fail in production because the model is bad. They fail because the system around the model has no guardrails, no structure, and no way to see what's happening.Most agents don't fail in production because the model is bad. They fail because the system around the model has no guardrails, no structure, and no way to see what's happening. In this session, I'll make the case that reliability is an engineering problem, not a prompting one. I'll walk through the harness every production agent needs: control flow, a real plan, execute, observe, and replan design, and observability you can act on. I'll back it with production failures we've all seen and a benchmark where the same model went from 20% to 60%+ just by changing the system. You'll leave with a pre-deploy checklist: the blocks to include, and the ones most teams forget.read more
Lorenze HayLead Open Source Engineer, crewAI
☆
Tue 9/29 · 11:00 AM - 12:30 PM
Go from idea to working product - without needing a full technical team.
● Day Zero: Leadership
Workshop
📍 Theater 2
Marily Nika walks you through a modern AI-powered product development process: validating demand, defining what to build, pressure-testing decisions with AI agents, creating prototypes, and evaluating whether the product actually works.Marily Nika walks you through a modern AI-powered product development process: validating demand, defining what to build, pressure-testing decisions with AI agents, creating prototypes, and evaluating whether the product actually works.
You’ll see where AI can replace work that once required additional research, design, engineering, and analysis support - and leave with a practical path from idea to working prototype and real user feedback.
What You’ll Learn1. Validate Real DemandUse AI-assisted research alongside real user conversations to uncover meaningful pain points, willingness to pay, and strong product signals before you build.
2. Turn Insights Into Focused RequirementsTranslate what you learn into a prioritized PRD that defines the user, core use cases, success metrics, and smallest valuable feature set.
3. Pressure-Test Your Product With AI AgentsCreate product, customer, engineering, and risk reviewers that challenge your assumptions around scope, feasibility, privacy, customer value, and overpromising.
4. Prototype From EvidenceTurn validated requirements into UI concepts, iterate quickly, and move the strongest ideas into the right no-code or AI-powered builder.
5. Know the Difference Between a Prototype and an MVPUse prototypes to gather evidence before investing further, then define MVP tests around desirability, feasibility, business viability, and your riskiest assumptions.
6. Measure Product QualityBuild rubrics and evaluation loops to assess AI outputs for relevance, groundedness, completeness, safety, and hallucination risk.read more
How AI-Powered Leaders Make Better Strategic Decisions
● Day Zero: Leadership
Workshop
📍 Theater 2
Leaders today face more data, more possibilities, and more uncertainty than ever. Yet making better strategic decisions is not simply a matter of moving faster or asking AI for an answer.Leaders today face more data, more possibilities, and more uncertainty than ever. Yet making better strategic decisions is not simply a matter of moving faster or asking AI for an answer. It requires knowing when to slow down, challenge your assumptions, and widen the lens.
In this interactive workshop, founders and executives will use practical tools from decision science, critical thinking, and leadership to work through a real strategic decision with or without AI as a thought partner. Bring a decision you are currently facing and leave with new perspectives, greater clarity, and a more thoughtful path forward.read more
Mo FongExecutive Coach, Adjunct Lecturer
☆
Tue 9/29 · 2:00 PM - 3:30 PM
Kill, Suspend, or Scale? A Hands-On AI Governance Workshop Using the CHOIR-G Framework
● Day Zero: Technical
Workshop
📍 Theater 3
The Problem: Standard technical metrics like Accuracy and Recall do not measure real-world risk, especially in high-stakes environments. This workshop moves beyond theoretical AI ethics to provide a hard, quantitative framework for deployment decisions.The Problem: Standard technical metrics like Accuracy and Recall do not measure real-world risk, especially in high-stakes environments.
This workshop moves beyond theoretical AI ethics to provide a hard, quantitative framework for deployment decisions. Participants will learn the CHOIR-G Framework, a risk-weighted scoring model, and apply it in real-time to a notorious healthcare case study, calculating risk to make a definitive "Kill, Suspend, or Scale" deployment decision.
As AI systems move from sandbox to production, engineering and leadership teams face a critical "AI Risk Gap." A model may boast high internal performance metrics, but if it fails in a live, dynamic environment, who is harmed? Who owns the risk outcomes? Are core error controls in place at reporting?
This 90-minute workshop equips AI builders and strategists with the CHOIR-G Framework: a vendor-neutral, highly actionable governance methodology that bridges technical performance with social trust and enterprise accountability. CHOIR evaluates five core dimensions: Context, Human Impact, Organizational Accountability, Indicators, and
Review & Oversight.
Instead of passively listening to a lecture, attendees will be put to work. After a brief breakdown of the CHOIR-G mathematics :
S = wC * C + wH * H + wO * O + wI * I + wR * R
where:
S = overall composite score
wC = weight assigned to the C dimension
wH = weight assigned to the H dimension
wO = weight assigned to the O dimension
wI = weight assigned to the I dimension
wR = weight assigned to the R dimension
C = score for the C dimension
H = score for the H dimension
O = score for the O dimension
I = score for the I dimension
R = score for the R dimension
The room will break into groups to act as a board for governing enterprise AI.
The Hands-On Case Study:
Groups will be tasked with conducting a governance review of the Epic Systems Sepsis Early Warning System. In internal testing, the model boasted a high AUC-ROC, but in production, University of Michigan researchers (published in JAMA Internal Medicine) found
it missed 67% of actual sepsis cases because the model "cheated" by picking up on when doctors were already ordering antibiotics.
Given the life-or-death stakes, teams must debate the weights of each CHOIR dimension for this specific deployment, score the system's failure, and determine the necessary governance outcome, ultimately working through the process to establish the strict "Human-in-the-Loop" protocols required to secure such a system.
What Attendees Will Walk Away Knowing:
• The CHOIR-G Scoring Mechanics: How to implement a 1 - 15 risk-weighted scoring system for any AI model.
• The "Critical 1" Automatic Intervention: How to establish non-negotiable guardrails in their deployment pipelines to prevent catastrophic edge-case failures.
• Cross-Functional Translation: How to translate abstract ethical and risk concerns into concrete, quantifiable metrics that engineering, legal, and executive teams can agree on.read more
Dr. Stella UmunnaAdjunct Professor, George Washington University
How to Build a Multimodal AI System That Runs Entirely on Your Own Hardware
● Day Zero: Technical
Workshop
📍 Theater 3
Most multimodal AI systems depend on cloud APIs which means your data leaves your device, your system breaks when the API goes down, and your costs scale with every request. This workshop cuts that dependency entirely.Most multimodal AI systems depend on cloud APIs which means your data leaves your device, your system breaks when the API goes down, and your costs scale with every request. This workshop cuts that dependency entirely. We build a fully local multimodal AI system from scratch that takes live camera input, processes audio, and returns intelligent responses all running on your own hardware, with no external API calls.
Attendees will leave with a working system they built themselves and a clear understanding of every architectural decision made along the way.
What we build:
A local AI assistant that takes live camera input and voice commands, runs inference on-device using a local multimodal LLM, and responds in real time no internet connection required.
Workshop arc:
1. Setup and architecture overview - local LLM options, hardware requirements, why on-device matters
2. Vision layer - wiring in camera input, frame preprocessing, feeding images to a local multimodal model
3. Voice layer - capturing audio input, running local speech-to-text, connecting it to the inference pipeline
4. Response layer - grounding responses in visual and audio context, handling multi-turn conversations
5. Stress testing, Q&A, and where to take it nextread more
Kavya Sri ChennojuStaff AI Engineer, Arm
☆
Tue 9/29 · 4:00 PM - 5:30 PM
Learnings Put into Practice: A real-world, scenario-based discussion for startups using AI models to provide & develop their products/services
● Day Zero: Leadership
Workshop
📍 Theater 2
This final session of the Leadership Track will bring together the day’s learnings and challenge you to apply them to a real-world scenario.This final session of the Leadership Track will bring together the day’s learnings and challenge you to apply them to a real-world scenario. You and your fellow session mates will need to find solutions for a startup facing an operational and leadership dilemma, giving you the opportunity to test-run your ideas in a safe environment. You will present your ideas and vigorously defend your positions, incorporating the new insights you acquired during Day Zerø as well as incorporating new concepts brought up that may have been unforeseen. You will need to react to actual situations facing startup leaders when incorporating AI use into their business models. You’ll hear different perspectives from the diverse audience and reinforce or reevaluate your thoughts as you might in a discussion with senior industry advisors. We will set the stage for the workshop with initial coverage of AI model cost considerations, AI model evaluation and selection, AI development and process transparency considerations, legal considerations (model use liability as well as client-produced generative solution liability), client and internal company data stewardship, as well as team acceptance and management during AI use for the products/services. This facilitator-led case discussion will require you to think on your feet, reacting in real-time to market impediments and business obstacles. You should be prepared for cold calls and bring your specific industry knowledge into the discussion, utilizing fact-based points to present your argument or perspective. You’ll leave this session armed with new ideas and well-discussed points you can apply immediately to your firm’s AI approach. The learnings from this workshop will be memorable from the engaged discussions, impactful from the robust debates, and actionable from the vetted deliberations.read more
John BlevinsProfessor, Cornell Tech, UC Berkeley School of Information, and UCLA Anderson
Built for you by(617)-PATRICK, Champion AI speaker
AI is moving from breakthrough to breakthrough - but where is it actually taking us?AI is moving from breakthrough to breakthrough - but where is it actually taking us? Peter Norvig looks beyond today's models and benchmarks to explore what the next era of AI could bring, which assumptions may need to change, and the questions that will matter most as intelligent systems become more capable and widespread.read more
Peter NorvigDistinguished Education Fellow, Stanford HAI
☆
Wed 9/30 · 9:25 AM - 9:40 AM
Beyond the Frontier: The economics of the multi-model era
● Keynotes
Keynote
📍 Theater 1
For the last few years, the AI conversation has been dominated by a handful of frontier, closed-weight models but that era is ending.For the last few years, the AI conversation has been dominated by a handful of frontier, closed-weight models but that era is ending. As open-source models have become better and better and new specialized model labs have emerged, the industry is shifting toward a multi-model future. This future is one where open source and closed-weight models are distributed, evaluated, and run side by side on shared infrastructure, and where the right model is chosen on quality and performance for a specific job rather than the lab who created it.
Baseten sits at the center of this shift, serving as a hub for distributing both open source and closed-weight models at production scale. In this talk, Philip Kiely, author of Inference Engineering and early employee at Baseten, will make the case for why this shift is not just inevitable but good for the industry. The unit economics of open source models now dramatically outperform frontier models for equivalent tasks, and smaller, fine-tuned models frequently outperform large general-purpose models on quality, latency, and cost for real customer workloads.
Drawing on data from production deployments, this talk will show why the future of AI isn't one model for all tasks. It's the right model, for the right job, at the right price.read more
Language Models, Inside Out: Finding the world models hidden in the weights
● Keynotes
Keynote
📍 Theater 1
Language models are superhuman text predictors. The best way to predict a piece of text is to build a rich model of the world that generated it.Language models are superhuman text predictors. The best way to predict a piece of text is to build a rich model of the world that generated it. Look inside a language model and that's what you find: concepts shared across languages and modalities, plans formed well in advance of being enacted, and even evolved algorithms for arithmetic. Prediction, done well enough, becomes simulation. To guess the next word, the model learns to represent the hidden variables behind the text: a speaker's emotions, beliefs, and goals.
The field of interpretability has spent the last few years decoding language model weights directly. This talk is a tour of some of the most important findings from this work, and of the mental model they add up to: a language model as a world model, hiding in plain sight in the weights.read more
Agentic Commerce - A new paradigm for financial, legal, and end-user interaction
● Keynotes
Keynote
📍 Theater 1
Existing payment rails, and all that is built around it, from customer identity verification, to terms of service, to the tools you use to manage and facilitate payments were built with the human user experience in mind.Existing payment rails, and all that is built around it, from customer identity verification, to terms of service, to the tools you use to manage and facilitate payments were built with the human user experience in mind. As companies and industries explore the use of AI in commerce, trading, and financial markets, they realize that the traditional systems used to carry out these transactions cannot be easily adapted for use by AI Agents. Companies from Stripe to Coinbase to Cloudflare are recognizing the new capabilities and technologies needed to enable agentic commerce and are building toward that future.
The way that you provide and control funds accessible to an agent, as well as how commercial flows like shopping and checkout are executed will need to be redesigned, as adapting human-centric rails doesn’t work in an AI-driven environment. Provisioning a virtual credit card for an agent doesn’t mean that agent will be able to use those funds in most commercial flows. This keynote will explore some of the gaps that exist in using traditional systems with AI, and propose some of the new solutions that can be adopted.
Enterprises will understandably need more control compared to a hobby builder to enable secure agentic workflows with commercial transactions, and this keynote will discuss what the future looks like for different parties. With a summary of what is being built and what standards are being developed, this talk will give a high level view of what a commerce-enabled agentic future looks like. Distributed Ledger Technologies (DLTs) are especially well suited to play a role in Agentic Commerce, providing the infrastructure for payments, solidifying and recording human intent, legal agreements, and audit trails for agentic actions.read more
AI hardware is exploding, but developers are still locked to a single programming model.AI hardware is exploding, but developers are still locked to a single programming model. Chris Lattner, creator of LLVM, Clang and Swift, co-founder and CEO of Modular,
and EVP of Advanced AI Software & Platforms at Qualcomm, argues the future of compute will be decided at the software layer, not the silicon. He introduces Mojo 1.0, now fully open source under Apache 2.0, and MAX,
an inference platform built to deliver portability and performance across increasingly diverse hardware without vendor lock-in or rewrites.read more
Chris LattnerCo-founder and CEO, Modular
☆
Wed 9/30 · 10:20 AM - 10:30 AM
GraphRAG: The Definitive Guide - A Field Guide to Patterns, Trade-offs, and Production Systems
● Keynotes
Keynote
📍 Theater 1
GraphRAG has quickly become a catch-all term for everything from graph-enhanced retrieval to Text2Cypher, agent memory, and community-based summarization. That is exciting, but it also creates confusion.GraphRAG has quickly become a catch-all term for everything from graph-enhanced retrieval to Text2Cypher, agent memory, and community-based summarization. That is exciting, but it also creates confusion. When people say “use GraphRAG,” what do they actually mean, and which pattern fits which problem?
This talk gives AI engineers a practical map of the GraphRAG landscape. We start with the limits of vector-only RAG, especially when data is connected, multi-hop, dynamic, or needs to be explainable. From there, we break GraphRAG into a set of core building blocks: constructing knowledge graphs from structured and unstructured data, choosing among lexical, domain, and meta graphs, and applying retrieval patterns such as guided traversal, graph-filtered search, Cypher templates, Text2Cypher, graph embeddings, query-focused summarization, and agentic retrieval.
The goal is not to present GraphRAG as one recipe, but as a toolbox with clear trade-offs. We will show when a lightweight pattern is enough, when a richer graph model pays off, and how to think about evaluation, governance, and production readiness from the start. Attendees will leave with a working mental model of GraphRAG, a taxonomy of the most useful patterns, and a decision framework for designing graph-powered GenAI systems that are accurate, explainable, and useful beyond the demo.read more
Stephen ChinVP, Developer Relations, Neo4j
Built for you by(617)-PATRICK, Champion AI speaker
AI is reshaping how companies are started, scaled, governed, and sustained. Tim O’Reilly and Eric Ries bring decades of perspective on technology cycles, founder responsibility, and the systems that determine which companies endure.AI is reshaping how companies are started, scaled, governed, and sustained.
Tim O’Reilly and Eric Ries bring decades of perspective on technology cycles, founder responsibility, and the systems that determine which companies endure. O’Reilly has helped shape some of the defining ideas of modern technology, from open source and Web 2.0 to mechanism design for the human-AI economy. Their conversation will cover Eric’s experience advising Anthropic, the ideas behind Incorruptible, and what it takes to build AI companies that matter.read more
Eric RiesFounder and Author of The Lean Startup and Incorruptible, LTSE Service
Beyond Toil: Building a Scalable AI-Powered Change Management Platform
● Applied AI
📍 Theater 4
This talk will share practical insights and design patterns for developing a scalable and efficient AI agent platform for generalized change management, applicable to Org Policy, IAM, code reviews, and beyond.This talk will share practical insights and design patterns for developing a scalable and efficient AI agent platform for generalized change management, applicable to Org Policy, IAM, code reviews, and beyond.
The session will cover:
- The "Synapse as a Service" Vision: Designing a multi-client platform where various teams can onboard their own review workflows.
- Dynamic Workflow Orchestration: Architecting a system that can intelligently route requests and dynamically load team-specific playbooks and logic.
- Intelligent Dispatching: Utilizing LLMs to understand and categorize incoming change requests from various sources (Bugs, Code changes, API calls).
- Playbook-Driven Evaluation: How natural language playbooks, maintained by client teams, guide the AI's review process, allowing domain experts to define logic without deep coding.
- Modular Action Execution: Designing a flexible system for taking actions based on review outcomes, such as updating tickets, commenting on CLs, or triggering other agents.
- RAG for Contextual Awareness: Strategies for Retrieval Augmented Generation (RAG) to provide the agent with the necessary context from internal knowledge bases and documentation.
- Key Design Considerations: Challenges and learnings in building a configurable and extensible multi-agent system using ADK, including state management and tool integration.
- Measuring Success: Defining metrics for evaluating platform effectiveness, including adoption, processing time, and consistency.
Future Roadmap: Expanding capabilities, onboarding more client teams, and deeper integrations.
This session provides a blueprint for leveraging AI to create adaptable and intelligent change management automation, drawing on our experience building the Synapse Agent Platform at Google.read more
Jack HopkinsSecurity Engineer, Google
Andrey KovalevSenior Security Engineer, Google
☆
Wed 9/30 · 11:30 AM - 11:55 AM
Evaluating Agentic AI in the Real World: A Framework for AI Evaluation Through the Use Case of a Retail Shopping Assistant
● AI Strategy
📍 Theater 5
Most teams treat evaluation as a checkpoint before launch. This talk argues it belongs earlier: baked into the development loop from the first prototype, not bolted on after the fact.Most teams treat evaluation as a checkpoint before launch. This talk argues it belongs earlier: baked into the development loop from the first prototype, not bolted on after the fact. It lays out an evaluation lifecycle for agentic AI systems, covering how eval requirements get gathered across stakeholders, how metrics get designed and validated, and how evals get embedded into CI/CD as a continuous gate rather than a one-time audit.
The framework is substantiated with SERA, a taxonomy of 21 LLM-as-a-Judge metrics spanning Safety, Compliance, and Agent Quality for retail agentic shopping assistants, accepted at RecSys 2026. The talk traces the evaluation lifecycle from metric design and risk-proportional aggregation to a persona-driven simulator for multi-turn stress testing, and ultimately to deploying the evaluation suite as a CI/CD gate for an agent under development. It highlights real regressions caught in practice including relaxed dietary filters, over-refusal, and degraded recommendation relevance while also examining the framework’s limitations and the challenges that remain in evaluating production-grade agentic systems.read more
Sowmya PodilaSenior Applied AI Scientist, Target
☆
Wed 9/30 · 11:30 AM - 11:55 AM
Monitoring Frontier Models for Frontier Risks
● AI Frontiers
📍 Theater 1
An OpenAI model breaking containment and attacking Hugging Face this summer proved that far-off frontier AI risks are now everyday AI risks.An OpenAI model breaking containment and attacking Hugging Face this summer proved that far-off frontier AI risks are now everyday AI risks. The industry is months away from an open model capable of the same damage, or worse, in higher-stakes emerging fields like biology. We need to prepare for these emerging risks by building guardrails that catch reward hacking, offensive cyberattacks, and biorisk in production, because these and other risks still slip past today's standard monitoring approaches. In this talk, Dan Balsam will present an approach to frontier risk guardrails that "reads the mind of the model," using a model's internal representations to catch reward hacking and other harmful behavior as it happens. He'll share case studies from production guardrails across risk categories like cyber and biology, discuss benefits for cost and performance, and what defenders can build into their models now to protect against frontier risks.
Bio: Dan Balsam is co-founder and CTO of Goodfire, an AI research lab using interpretability to understand, learn from and design AI systems.read more
Dan BalsamCo-founder and CTO, Goodfire
☆
Wed 9/30 · 11:30 AM - 11:55 AM
Stop Renting Intelligence: The Train-to-Deploy Loop for Specialized AI
● The AI Stack
📍 Theater 3
The next wave of AI products will not rely only on prompting generic frontier models. Winning teams will own specialized models shaped by their product data, user feedback, and domain workflows.The next wave of AI products will not rely only on prompting generic frontier models. Winning teams will own specialized models shaped by their product data, user feedback, and domain workflows.
In this 25-minute session, we’ll cover the practical loop behind model ownership: choose a base model, prepare data, fine-tune with SFT/DPO/RL, evaluate outputs, deploy the tuned model, collect feedback, and repeat. We’ll also explain why training and inference should be treated as one system, not separate steps.
Attendees will leave with a simple framework for when to tune, when RL matters, and how continuous improvement turns fine-tuning from a one-off project into a product advantage.read more
Jetashree RaviTechnical Lead Manager, Fireworks
☆
Wed 9/30 · 11:30 AM - 11:55 AM
Your agents' traces are your most valuable asset: reading the software factory floor
● AI Builders
📍 Theater 2
A modern harness does not click buttons. It writes a shell line, a pipe, an inline script, and hands it to an interpreter.A modern harness does not click buttons. It writes a shell line, a pipe, an inline script, and hands it to an interpreter. It does that in step with one inference backend, the frontier model it was trained against and benchmarked with. Swap the harness and you have not measured efficiency; you have changed the model, the prompt cache, the tool vocabulary and the compaction rules at once.
Olki sits beside the harness you already use - Codex and OpenCode - and records what actually ran. Each line becomes typed events, one per stage, with the program, its arguments, and what it touched. From that record, four readers get their own view: the organization across machines, the inference provider counting repeated failures and the cache that carries the bill, and the product owner asking how agents use the tool they shipped.
This session walks the software factory floor: how your harness interacts with the inference provider stack, how a prefix cache lives, why turn two must reach turn one's node, what a subagent or a model switch costs, and what you can only learn by reading the record instead of swapping the harness.read more
Behind One Useful Answer: Building Uber’s Earner Assistant
● Applied AI
📍 Theater 4
When a driver asks, “Why am I not getting trips?”, a fluent answer isn’t enough. A useful response must connect their situation with current marketplace conditions, explain relevant options, and leave the choice with the driver.When a driver asks, “Why am I not getting trips?”, a fluent answer isn’t enough. A useful response must connect their situation with current marketplace conditions, explain relevant options, and leave the choice with the driver. This talk shares lessons from building Uber Assistant for Drivers, a production AI system that helps drivers and couriers explore opportunities, understand earnings, and navigate everyday tasks.
We’ll cover the product and engineering decisions behind the answer: bringing together real-time, contextual data, designing tools around users’ tasks rather than backend service boundaries, and balancing reasoning depth and decision quality with latency and cost. We’ll also discuss centralized input and output guardrails for safety, communication policies, controversial topics, and adversarial requests. Through production examples, we’ll show how answers can sound plausible while missing relevant data or responding without appropriate empathy. Our evaluation approach combines LLM judges with human review, including the challenge of aligning human judgments to calibrate automated evaluators.
Finally, we’ll explore proactive in-app guidance, our work on voice, and the ambition guiding these choices: an agent that understands drivers’ goals and constraints and works on their behalf while keeping them in control. Attendees will leave with practical lessons for building AI products where accuracy, responsiveness, and trust matter.read more
Dharmin ParikhDir of Product, Uber
Himaanshu GuptaDir of Engineering, Uber
☆
Wed 9/30 · 12:00 PM - 12:25 PM
Controlling AI with AI: How to Secure Agentic AI at Scale
● AI Strategy
📍 Theater 5
AI agents operate at 100x human speed, completely outpacing traditional frameworks for enterprise oversight and control. This unprecedented speed brings unpredictable risk.AI agents operate at 100x human speed, completely outpacing traditional frameworks for enterprise oversight and control. This unprecedented speed brings unpredictable risk. Whether an agent accidentally goes rogue or is deliberately hijacked by a threat actor, the fallout is instantaneous."
In this session we'll confront a hard truth of the agentic enterprise: static, rules-based security and "human-in-the-loop" safeguards are dead. They simply cannot scale against unpredictable, non-deterministic systems. To unleash agents without the risk, you need a fundamentally new architecture where AI governs AI. Join us for a deep dive into this new approach and see exactly how it operates in the wild.
What you'll learn:
• Why legacy security fails against autonomous agents and what to do about it
• Why you need small language models to enforce semantic policies, instantly decoding intent where rigid machine rules break down
• Real world Use Case: How we secured Claude, proactively catching issues like data exposure and credential harvesting
• Live Tool Blocking Demo: Instantly identify and neutralize an unauthorized autonomous action before it happensread more
Dev RishiGM of AI, Rubrik
☆
Wed 9/30 · 12:00 PM - 12:25 PM
Data Quality is The Compute Multiplier
● The AI Stack
📍 Theater 3
Better data quality is the highest-leverage and most underinvested part of building a model: it produces a better model for the same compute, whether you're mid-training on an open base or pre-training from scratch.Better data quality is the highest-leverage and most underinvested part of building a model: it produces a better model for the same compute, whether you're mid-training on an open base or pre-training from scratch. This session is a practical look at data curation, covering what data quality actually means, the stages of a modern curation pipeline (cleaning, filtering, deduplication, synthetic data generation, algorithmic mixing, and multi-stage composition), and which steps matter most in practice. It draws on DatologyAI's frontier data research and customer results, including Thomson Reuters' mid-training gains on proprietary legal domain data and Arcee's Trinity model reaching the open frontier on public data alone. You'll leave with a concrete sense of where better data quality pays off and how data curation is shaping the future of model training.read more
Ari MorcosCo-Founder/CEO, DatologyAI
☆
Wed 9/30 · 12:00 PM - 12:25 PM
Enterprise AI Engineering: From Modernization to Autonomous Software Delivery
● AI Frontiers
📍 Theater 1
Agentic software engineering is expanding beyond coding assistance toward autonomous execution across the software lifecycle.Agentic software engineering is expanding beyond coding assistance toward autonomous execution across the software lifecycle. As this shift takes hold, AI agents are beginning to take on broader, longer-running work, while humans increasingly define goals, constraints, and desired outcomes rather than individual tasks.
Application modernization illustrates this shift in practice: agents help understand complex systems, transform applications, and optimize how software is built and maintained. As agents take on this kind of work, enterprise realities come into play - hybrid and on-prem environments, governance, security, and operational control all shape how these systems can operate responsibly at scale.
As agents operate within these enterprise constraints, their workloads grow, and economics and model strategy become part of the architecture. This session explores how model selection and rotation can optimize for capability, speed, and cost, and what these trends suggest about the future of increasingly autonomous software delivery.read more
Anthony StevensPrincipal AI Architect, IBM
☆
Wed 9/30 · 12:00 PM - 12:25 PM
Your Agent Needs a Steering Wheel
● AI Builders
📍 Theater 2
What if your AI agent could recognize a bad move before making it and adjust its next step? Most agents rely on prompts for instructions and guardrails to catch failures.What if your AI agent could recognize a bad move before making it and adjust its next step? Most agents rely on prompts for instructions and guardrails to catch failures. That leaves a gap during execution, when an agent skips a required check, repeats a broken tool call, or reaches a decision that needs human judgment.
This session introduces steering, for just-in-time guidance inside the agent loop. Through a live demo using an open-source agent harness framework, attendees will learn how to inspect the proposed actions, guide recovery, enforce required steps, and add approval points without giving up the flexibility that makes agents useful.read more
Sandhya SubramaniSenior Gen AI Developer Advocate, Amazon Web Services
Built for you by(617)-PATRICK, Champion AI speaker
Fast Track to Production: The New Stack for Today's Founders
● The AI Stack
📍 Theater 3
Watch a complete software delivery pipeline go from zero to running in minutes, then spend the rest of the session on the more interesting question: what do you do with all the time you just got back?Watch a complete software delivery pipeline go from zero to running in minutes, then spend the rest of the session on the more interesting question: what do you do with all the time you just got back?
CircleCI's Nick Smith will demo the new CircleCI CLI live, and and lead a discussion on how tools with robust terminal-based access reduce the burden of setup so you can get to building software. Whether you're a technical founder, a domain expert eyeing your first product, or an executive wondering why your team still struggles with deployments, this session shows how much simpler getting code into production can be today, and opens the floor to what that possibility actually unlocks.read more
Nick SmithStaff Field Engineer, CircleCI
☆
Wed 9/30 · 12:30 PM - 12:40 PM
How to build an AI-native company: What Shopify, Instacart, Ramp, and Gusto do differently
● Lunch Sessions
Mainstage Talk
📍 Theater 1
After two years working with companies like Shopify, Instacart, Ramp, and Gusto, Max Brodeur-Urbas, CEO of Gumloop, has found that the hardest part of making AI work has almost nothing to do with the latest frontier lab model.After two years working with companies like Shopify, Instacart, Ramp, and Gusto, Max Brodeur-Urbas, CEO of Gumloop, has found that the hardest part of making AI work has almost nothing to do with the latest frontier lab model. It has to do with how the company is wired.
Nearly every company is in a sprint to avoid falling behind on the AI curve, rapidly rolling out agents and automations without considering who is best equipped to build them, how they will be governed, or what they will cost long term.
The organizations pulling ahead let the people closest to the work build and share automations across their team, they give IT the controls to govern AI usage, and they prioritize sustainable spending habits by routing work to cheaper open-weight models.
Max will walk through what this looks like inside companies already running this way, and leave leaders with a clearer picture of how to roll out AI, measure success, and see real business impact.read more
Max Brodeur-UrbasCEO, Gumloop
☆
Wed 9/30 · 12:30 PM - 12:55 PM
Observe, Enforce, and Govern: The Fiddler Control Plane and Centor Models
● AI Builders
📍 Theater 2
An agent in your pipeline just wrote code, called an API, or touched a database, and nobody reviewed it first. That's the autonomy agents already have.An agent in your pipeline just wrote code, called an API, or touched a database, and nobody reviewed it first. That's the autonomy agents already have. The question is whether you'd know if one did something wrong, and how long it would take you to find out.
That’s what the AI Control Plane is for: a system of trust that starts with evaluation, since that's what makes it possible to monitor an agent's activity, enforce policy on an agent's actions before they execute, and govern behavior across the agentic lifecycle.
We'll show what this looks like in practice: inline guardrails that check an agent's actions before they execute, and dashboards that break down exactly what every agent is doing and spending by developer, team, and token. We'll also show how Fiddler Centor Models come batteries-included, no external API calls, no added cost per evaluation, so you can evaluate 100% of your traces or scale down to sampling for smaller deployments. The same approach works whether you're operating a handful of agents or a fleet.
You'll leave with a practical way to decide which of your agents have earned more autonomy, and how to give it to them without losing the ability to govern what they're doing.read more
Joshua RubinPrincipal AI Scientist, Fiddler AI
☆
Wed 9/30 · 12:30 PM - 12:55 PM
The Instruction Layer: Turning Enterprise Knowledge Into Trusted AI Answers and Actions
● Lunch Sessions
Breakout Session
📍 Theater 4
Every AI-powered interaction now speaks for the company that deployed it. Users do not distinguish between an answer from a person and an answer from an agent, and neither do regulators.Every AI-powered interaction now speaks for the company that deployed it. Users do not distinguish between an answer from a person and an answer from an agent, and neither do regulators. Yet most teams have pointed capable models at operational knowledge that is fragmented, out of date, and ungoverned, then been surprised when benchmark scores hold up and production does not.
Inconsistent policies, duplicate content, conflicting sources of truth, and expertise that lives only in people’s heads produce the same outcomes every time: wrong answers, failed automation, compliance exposure, and a slow erosion of user trust that is expensive to win back.
This session looks at how leading organizations are building a trusted knowledge foundation for enterprise agents, moving from traditional knowledge management to infrastructure designed to instruct AI. We will cover what changes in governance, how to evaluate answer quality continuously rather than by sample, and what it takes to deliver instant, trusted answers across every channel an agent touches.read more
James HuntVP, Product Management & Solution Strategy, eGain
☆
Wed 9/30 · 12:30 PM - 12:55 PM
Your agents lack context: Here's how to fix "You're absolutely right!"
● Lunch Sessions
Breakout Session
📍 Theater 5
Agents can generate code. Very few can generate the right code for your organization, because they’re missing context.Agents can generate code. Very few can generate the right code for your organization, because they’re missing context. They don’t know why your team chose Redis over DynamoDB, what the team decided in a Slack thread from two months ago about the auth migration, or which architectural patterns your principal engineers actually enforce in review.
This talk is a practitioner’s guide to building a context engine: the reasoning layer that continuously synthesizes organizational knowledge across disparate sources into unified, queryable understanding. I’ll walk through the problems you actually have to solve - reasoning across systems that don’t agree with each other, searching globally before you can reason, maintaining identity-scoped permissions so every user and agent only sees what they should, and personalizing results based on who’s asking and what they’re working on. These are the engineering challenges that make naive RAG fall short, drawn from real lessons building this at scale.read more
Brandon WaselnukDeveloper Relations, Unblocked
☆
Wed 9/30 · 12:45 PM - 12:55 PM
Introducing Redpanda's Agentic Data Plane
● Lunch Sessions
Lightning Talk
📍 Theater 3
AI agents are increasingly deployed as digital employees - accessing enterprise data, making decisions, and acting autonomously.AI agents are increasingly deployed as digital employees - accessing enterprise data, making decisions, and acting autonomously. But agents are both less predictable than humans (prone to hallucination, misinterpretation, and adversarial manipulation) and more capable, with deep system access and machine-speed execution that can cascade damage fast. This makes it unsafe to trust agents to faithfully interpret or propagate security-critical metadata like access policies, data classifications, and behavioral constraints.
We present Redpanda's Agentic Data Plane (ADP): an architecture using out-of-band metadata channels that carry security context, policy signals, and audit trails deterministically - entirely outside the agent's read/write path, across heterogeneous infrastructure. These channels enforce governance at every stage of the agent lifecycle: scoping data access going in, constraining actions during execution, and capturing tamper-proof audit trails coming out.read more
Piyush AgarwalVP, Engineering, Redpanda
☆
Wed 9/30 · 12:45 PM - 12:55 PM
Move Fast and Don’t Break Things: A New Model for AI Governance
● Lunch Sessions
Mainstage Talk
📍 Theater 1
Every enterprise wants to move faster with AI, but most governance programs slow that down because they were built for humans, not machines.Every enterprise wants to move faster with AI, but most governance programs slow that down because they were built for humans, not machines. AI agents act in milliseconds, often with more access than any single task requires, and most governance programs lose sight of what they’re doing once they’re running.
Companies often feel forced into a false choice: lock AI down and slow the business or move fast and lose control. The companies pulling ahead rejected that trade-off entirely. They rebuilt their controls to match the speed of AI, getting more agents into production faster, with a complete record of every action to back it up.
This session shows how, and what becomes possible when authorization stops being a policy you set once and starts keeping pace with every decision your AI makes.read more
Art GillilandCEO, Delinea
Built for you by(617)-PATRICK, Champion AI speaker
Controlling the AI Software Factory: Keeping Engineers in Charge at Agent Speed
● Lunch Sessions
Breakout Session
📍 Theater 2
AI is not just changing how fast software is written. It is changing how software is produced.AI is not just changing how fast software is written. It is changing how software is produced.
Engineering organizations are assembling agent-powered systems that write code, review pull requests, validate releases, deploy changes, and continuously improve production software around the clock, without human hands on every step. The promise of faster software delivery and continuous optimization is real. But as delivery becomes more autonomous, a new and more urgent problem surfaces: how do you stay in control of systems that never stop evolving?
In this session, Jonathan Nolen, Chief Product Officer at LaunchDarkly, makes the case that the defining challenge of the AI era is not generating code - it is controlling autonomous systems in production. Traditional pre-production controls and the standard playbook of detect, fix, redeploy break down for agents that don't fail like traditional software. They drift. You can no longer catch issues before production. You have to control production.
Jonathan will share what LaunchDarkly has learned building its own AI software factory - including what happened when coding agents met its oldest, most complex production code - and why every autonomous delivery system requires a closed control loop: observation, decision-making, intervention, and feedback, all running at runtime without a deploy cycle.
Attendees will leave with a practical framework for building AI software factories that are genuinely fast and safe, and a clear answer to the question every engineering leader is now facing: when agents are doing the work, who stays in charge?read more
Jonathan NolenChief Product Officer, LaunchDarkly
Timothy CookDir of Product Management, LaunchDarkly
☆
Wed 9/30 · 1:00 PM - 1:25 PM
Mission control - Giving AI agency (without losing control)
● Lunch Sessions
Breakout Session
📍 Theater 5
The AI conversation has become uncomfortably binary: either hand everything to AI and hope for the best, or resist automation entirely and fall behind.The AI conversation has become uncomfortably binary: either hand everything to AI and hope for the best, or resist automation entirely and fall behind. But organizations achieving real results are doing something different: building Mission Control architectures where AI has genuine agency while humans maintain strategic oversight. Inspired by NASA’s model of autonomous systems backed by human judgment, this session will show you how to determine when AI should act independently, when it should assist, and when people must remain in control. You’ll leave with a practical framework for scaling AI across service operations without sacrificing visibility, trust, or control.read more
Monitoring in the Age of GenAI: Beyond Golden Signals
● Lunch Sessions
Lightning Talk
📍 Theater 3
The four golden signals (Latency, Errors, Traffic, Saturation) have been the foundation of application monitoring for years, and it still matters, but for GenAI applications, these signals alone leave significant blind spots.The four golden signals (Latency, Errors, Traffic, Saturation) have been the foundation of application monitoring for years, and it still matters, but for GenAI applications, these signals alone leave significant blind spots. A request can return 200 OK with low latency while the response hallucinates, leaks PII, or costs much more than expected.
This talk will walk you through what changes when you're monitoring non-deterministic, token-priced, prompt-injectable systems. We'll cover three additional monitoring dimensions - Cost (token attribution, model-mix tracking, wasted spend on failed requests), Safety (prompt injection detection, PII scanning, jailbreak attempts), and Quality (hallucination rate, relevance scoring, user satisfaction) - and show why each one is necessary alongside your existing instrumentation.read more
Lars KampSenior Staff Technical Advocate, Datadog
☆
Wed 9/30 · 1:00 PM - 1:10 PM
Secure Orchestration Layer for Agentic Enterprise
● Lunch Sessions
Mainstage Talk
📍 Theater 1
As AI models and agents move into production, inference optimization and security are becoming critical infrastructure challenges for enterprises.As AI models and agents move into production, inference optimization and security are becoming critical infrastructure challenges for enterprises.
This talk will explore how inference can be optimized at the token level to improve latency, and cost efficiency, particularly as agentic workloads become more demanding.
It will also address secure inference architectures and the technical requirements for protecting sensitive data throughout execution, while maintaining the performance required for production workloads.
Attendees will gain a practical understanding of how token-level optimization and secure inference can work together to run AI efficiently and securely at enterprise scale.read more
Richard RizkCTO, Blackbox AI
☆
Wed 9/30 · 1:00 PM - 1:25 PM
When AI Adoption Works: The Infrastructure and People Problems That Come Next
● Lunch Sessions
Breakout Session
📍 Theater 4
In this talk, Coder Product Manager David Fraley explores what happens after enterprise AI adoption gains traction and why success creates a harder set of problems than the ones platform teams started with.In this talk, Coder Product Manager David Fraley explores what happens after enterprise AI adoption gains traction and why success creates a harder set of problems than the ones platform teams started with.
Few organizations plan for what breaks once early barriers are cleared: code generation outpacing CI/CD, team structures misaligned with agent workflows, and senior engineers shifting from writing code to directing agents.
Drawing on patterns across enterprises at every stage of AI maturity, this session maps predictable failure modes at scale and the infrastructure decisions that support or stall the transition.
Key takeaways:
* Why CI/CD becomes the bottleneck after AI rollout
* How team structure must evolve as agents scale
* What sustaining AI adoption looks like in practiceread more
David FraleyProduct Manager, Coder
☆
Wed 9/30 · 1:15 PM - 1:25 PM
Actions speak louder than prompts: governing what AI agents actually do
● Lunch Sessions
Mainstage Talk
📍 Theater 1
Everyone is watching what goes into AI agents. Almost no one is watching what they actually do.Everyone is watching what goes into AI agents. Almost no one is watching what they actually do.
The industry obsesses over the inbound boundary - prompt injection, guardrails, model alignment - while the real risk lands on the other side: the commands an agent runs, the data it moves, the credentials it reaches for, the systems it changes. An agent doesn't need to be jailbroken to cause damage; it just needs to be helpful in the wrong direction.
Existing controls can't see this. Every tool in the security stack - secrets managers, DLP, identity, endpoint - watches a boundary: a vault checkout, a network egress, a login. Agents do their work, and their damage, in the space between those boundaries, where every action looks routine and the only identity in sight is a trusted human's. Prompts are too fuzzy to be controls; audit trails are precise but after the fact. Neither answers the question that matters: should this action happen, right now, in this context?
This talk makes the case for a new layer: understanding each action an agent is about to take - what will run, what it will touch, and why - and evaluating it before it executes. I'll share real patterns of risky agent behavior from enterprise environments that are invisible to today's tooling.
Whether you're enabling AI or governing it, you'll leave with a practical framework for the action layer - and the questions your current stack can't answer.read more
Nathan WallaceFounder/CEO, Turbot
☆
Wed 9/30 · 1:15 PM - 1:25 PM
Why the AI You Love Can't Touch Your Data (Yet)
● Lunch Sessions
Lightning Talk
📍 Theater 3
You've felt it. You sit down with a frontier AI and watch it do something genuinely brilliant. So why isn't that capability showing up in your work?You've felt it. You sit down with a frontier AI and watch it do something genuinely brilliant. So why isn't that capability showing up in your work? Two reasons: you can't hand company and customer data to a chat window, and your business isn't a chat. It's thousands of repeatable, auditable interactions a day, and nothing about raw model access makes those reliable, private, or provable. That gap is why an estimated 95% of enterprise AI pilots never reach production. At SS&C, we hit this wall at scale, processing billions of tokens a month across a regulated, global business, and had to build the layer that closes it: guardrails, monitoring, and auditing between the model and the work. In ten minutes, we'll show you what that layer looks like, why every organization needs one, and what changes when it exists.read more
Dave SergeantSenior Solutions Consultant, SS&C Blue Prism
Built for you by(617)-PATRICK, Champion AI speaker
Five lessons from building an open-source managed agents runtime
● Lunch Sessions
Lightning Talk
📍 Theater 3
Most agent tooling optimizes the agent loop. In production, that is almost never what breaks. We built a managed agents runtime - the layer between an agent framework and the infrastructure it actually runs on.Most agent tooling optimizes the agent loop. In production, that is almost never what breaks.
We built a managed agents runtime - the layer between an agent framework and the infrastructure it actually runs on. Five lessons from that, across the three problems nobody demos: building agents that outlive your framework choice, running them so they survive failure, and governing what they can spend, touch, and see.read more
Sijie GuoCEO, StreamNative
☆
Wed 9/30 · 1:30 PM - 1:40 PM
From AI-Assisted to AI-Native in 12 Months
● Lunch Sessions
Mainstage Talk
📍 Theater 1
What does it actually take to go all-in on AI as an engineering organization?What does it actually take to go all-in on AI as an engineering organization? This session traces monday.com's 12-month journey through three levels of AI engineering: developers pair-programming with AI, builders working in agentic mode, and finally autonomous agents owning complete tasks end-to-end. Anya shares the real metrics behind the shift - from +14% developer throughput at Level 1 to autonomous agents now owning 45% of all merged PRs, with overall output more than doubling. You'll see what this looks like in practice, with agents operating as genuine team members across code review, incident triage, and performance analysis. And you'll hear where monday.com is headed next: optimizing the human-agent team as a single unit.read more
Anya TchernishovEngineer, Monday.com
☆
Wed 9/30 · 1:30 PM - 1:55 PM
Moving Confidential Computing up the Stack: Taking AI Security from Theory into Practice
● Lunch Sessions
Breakout Session
📍 Theater 5
Confidential computing has reached AI, but mostly as a point solution for inference. As technology matures, each layer of the AI lifecycle - training, fine-tuning, reinforcement learning, and multi-party workloads - needs attested enclaves at scale.Confidential computing has reached AI, but mostly as a point solution for inference. As technology matures, each layer of the AI lifecycle - training, fine-tuning, reinforcement learning, and multi-party workloads - needs attested enclaves at scale. We present a blueprint and a reference implementation: Kubernetes-orchestrated Confidential Containers, auto-scaling attested GPU enclaves, and an attestation-bound key exchange that keeps sole key custody with the data owner.
Three Key Takeaways
1. Confidential computing is the foundation of data sovereignty. Moving beyond just inference, all AI training and inference workloads require zero-trust built directly into the compute level. This ensures that data in use is fully accounted for in security, extending protection across the entire ML lifecycle - including pre-training, fine-tuning, and multi-party training on regulated corpora - so that every stage relies on attested, verifiable workloads.
2. Standard deployment workflows must be supported. Manual, per-node attestation by engineers is neither scalable nor how enterprise customers manage their workloads. Deployment workflows must remain consistent with existing standards, while being augmented to support attested TEE deployment at scale. Using Kubernetes with Confidential Containers allows existing deployments to be retrofitted with minimal effort but still providing a trusted way of attesting at scale.
3. Scalability must not compromise data sovereignty. While high-scale deployment is essential, it shouldn’t require delegating key management to a third party. To maintain the integrity of trusted enclaves, the data owner must retain sole custody of encryption keys throughout the lifecycle. True confidential infrastructure achieves scale by offloading operations to the platform, while the owner controls the key release chain from end to end, ensuring the platform only ever handles ciphertext.read more
Sachin NeneHead of R&D, Corvex
☆
Wed 9/30 · 1:30 PM - 1:55 PM
What We Learned by Analyzing 1M AI-Generated PRs
● Lunch Sessions
Breakout Session
📍 Theater 2
Background coding agents are quickly moving from novelty to real-world software development workflows.Background coding agents are quickly moving from novelty to real-world software development workflows. Based on Greptile's analysis of millions of pull requests across 65,000 organizations, this talk explores how often end-to-end AI-generated PRs are being used and how their quality compares to human-written code. The data shows detectable agent-generated PRs grew from under 1% in February 2025 to 27.6% in April 2026, with early quality signals like revert rates and code churn suggesting these agents may already be competitive in serious codebases.read more
Daksh GuptaCo-founder and CEO, Greptile
☆
Wed 9/30 · 1:30 PM - 1:55 PM
What most AI gateways and goldfish have in common
● Lunch Sessions
Breakout Session
📍 Theater 4
Enterprise AI leaders and analysts now largely agree that data connectivity and context drive agent trust and accuracy. Catalogs, ontologies, semantic layers, and context repositories put meaning in a layer beside the data, and let the agent ask it.Enterprise AI leaders and analysts now largely agree that data connectivity and context drive agent trust and accuracy. Catalogs, ontologies, semantic layers, and context repositories put meaning in a layer beside the data, and let the agent ask it. The question now has become: should this layer sit next to or within your AI gateway infrastructure?
A layer outside the request path introduces two challenges. First, it delegates all policy and permission enforcement to the gateway itself, often resulting in broad-based data access. Second, it can only answer a question with the user-provided context, and like a goldfish it forgets the question and the context as soon as it routs it.
A layer inside the gateway sees and learns from the question, the cross-source relationships, the data that came back, and whether the answer survived contact with the user. It enables users and agents to confidently run both deterministic AI workflows and non-deterministic analytical queries, while unlocking a host of read-write use cases that have previously been shelved for security concerns.
This breakout session explores the benefits and tradeoffs of this critical architectural decision, as well as the cross-system gateway capabilities that not only manage but actively reduce token spend, drive agent accuracy, and unlock previously shelved use cases across enterprise deployments.read more
How to win the battle for the user in the agentic age
● Lunch Sessions
Lightning Talk
📍 Theater 3
The AI era looks like a fight over models, chips, and compute. The real fight is over the user. Whoever delivers the best experience wins the market, and agents are now the front line of that experience.The AI era looks like a fight over models, chips, and compute. The real fight is over the user. Whoever delivers the best experience wins the market, and agents are now the front line of that experience.
Most agents are losing it. They launch blind to the people they serve and miss the moments that decide whether a user stays or leaves. The gap between shipping an agent and shipping one people actually rely on is product context: the behavioral and business signals that tell an agent when to step in, what to say, and whether it worked.
This talk makes the case that the companies who win the age of AI won't be the ones with the most agents. They'll be the ones whose agents truly understand the user, and shows what it takes to build them.read more
Arif PoonawalaVP, Product Management, Pendo
☆
Wed 9/30 · 1:45 PM - 1:55 PM
No Agentic Sovereignty, No Agentic Security
● Lunch Sessions
Mainstage Talk
📍 Theater 1
For high-security organizations like government, finance, defense, and regulated industries, failing to secure agents isn’t just inconvenient. It’s catastrophic: stalled payments, delayed medical treatments, and dead power grids.For high-security organizations like government, finance, defense, and regulated industries, failing to secure agents isn’t just inconvenient. It’s catastrophic: stalled payments, delayed medical treatments, and dead power grids. More than 70% of regulated-industry leaders name security and privacy as their top barrier to adoption.
Agents break the assumptions these industries were built on. Unlike deterministic software, they take actions on their own, reach into live systems, and act with privilege no one formally provisioned. The autonomy is the value. Where it acts, and on whose authority, is the exposure.
The instinct is to bolt on data residency and call it solved. It isn't. Data sitting in a database is one thing; data an agent picks up, reasons over, and acts on is another - and that motion is where sovereignty actually breaks.
In this session, we’ll detail why regulated industries need truly sovereign solutions to support agents in the environments where they run. We'll examine why data residency alone isn't enough, and unpack the risks across every sovereignty control point an agent touches - data, infrastructure, decision, and ecosystem
Attendees will leave understanding the high-risk actions agents can take on their behalf, why probabilistic guardrails satisfy none of these dimensions, and what high-assurance controls - spanning identity, infrastructure, and decision authority - regulated industries actually need before they let agents act.
Organizations that treat this as a continuous, auditable posture will be the ones who move fast in AI adoption and still pass the audit.read more
Nandini V.Principal Product Manager, RSA Security
Built for you by(617)-PATRICK, Champion AI speaker
Context Engineering for AI Agents: A Practical Architecture for Production Systems
● AI Builders
📍 Theater 2
As AI agents move from demos into production, many teams discover that a more powerful model does not automatically produce better results.As AI agents move from demos into production, many teams discover that a more powerful model does not automatically produce better results. The real challenge is often context: giving the agent the right information, at the right time, in the right form.
This session shares lessons learned from building and scaling Incident Manager Assist, an enterprise multi-agent system designed to support complex incident investigations. As adoption grew and more enterprise data was connected, we encountered recurring problems: too much information overwhelmed the model, long-running investigations lost important decisions, answers were not tailored to the user, operational data became stale, and expensive models were used even for simple tasks.
These challenges led us to six practical context engineering patterns: model-driven control flow, scoped context assembly, intent-driven layered memory, user-aware context, task-adaptive model routing, and continuous context.
Attendees will leave with a clear way to think about context, and practical patterns for making context-aware agents reliable in production.read more
Bhavin ParikhSenior Principal Architect, Comcast
☆
Wed 9/30 · 2:00 PM - 2:25 PM
From LLMs to Language Agents: Reinforcement Learning in the Post-Training Era
● The AI Stack
📍 Theater 3
Pretraining builds knowledge. Reinforcement learning builds behavior. As LLMs move into production systems powering copilots, decision engines, and autonomous workflows post-training with RL has emerged as a critical step in shaping model behavior beyond imitation.Pretraining builds knowledge. Reinforcement learning builds behavior.
As LLMs move into production systems powering copilots, decision engines, and autonomous workflows post-training with RL has emerged as a critical step in shaping model behavior beyond imitation. This talk presents a practical guide to post-training language agents using reinforcement learning, covering reward modeling, human preference optimization, environment-based feedback, and tool-integrated training loops.
The talk will analyze trade-offs between supervised fine-tuning and RL, discuss stability and scaling challenges, and present real-world examples where RL improves reasoning depth, tool reliability, and policy alignment.read more
Rucha ApteGenAI Solutions Architect, NVIDIA
Annie SurlaProduct Research Engineering Manager, Deep Learning, NVIDIA
☆
Wed 9/30 · 2:00 PM - 2:25 PM
One Deployment, Five Capabilities: Building a Production Voice Agent on Edge Compute
● AI Strategy
📍 Theater 5
Building a voice agent that answers calls, sends follow-up SMS, and emails a transcript is easy in a hackathon. Shipping it to production with sub-200ms round-trip latency, carrier-grade call quality, and zero vendor finger-pointing is where most projects die.Building a voice agent that answers calls, sends follow-up SMS, and emails a transcript is easy in a hackathon. Shipping it to production with sub-200ms round-trip latency, carrier-grade call quality, and zero vendor finger-pointing is where most projects die.
This session walks through taking a voice AI agent from prototype to production on a single platform. The demo deploys a site on edge compute that handles inbound calls with an open-source model, follows up over SMS and email, and runs the entire stack from one CLI command. The talk covers the architecture decisions that matter at production scale: where your model runs relative to your telephony termination, how to handle stateful conversations without a separate orchestration vendor, and why the number of vendors in your stack is the strongest predictor of whether your agent survives contact with real traffic.
Attendees will see a working deployment built with the Telnyx CLI, learn the framework for evaluating stack complexity, and leave with a repo they can fork.read more
Stephen MalitoDir of Solutions Engineering, Telnyx
☆
Wed 9/30 · 2:00 PM - 2:25 PM
Why Are We Still Governing AI Like Software?
● AI Frontiers
📍 Theater 1
The biggest mistake we may be making with AI agents is treating them like software. Organizations have faced this problem before: how do you delegate real authority without giving up control?The biggest mistake we may be making with AI agents is treating them like software.
Organizations have faced this problem before: how do you delegate real authority without giving up control?
The answer may already exist in a place few AI teams are looking.read more
Tim TylerHead of Industry Engagement, Gravitee
☆
Wed 9/30 · 2:00 PM - 2:25 PM
Your Evals Probably Aren’t Dev-Friendly
● Applied AI
📍 Theater 4
Evals are only useful if developers can understand what changed, why it changed and what to do next.Evals are only useful if developers can understand what changed, why it changed and what to do next. Most eval systems reduce AI behavior to a single score, leaving engineers to dig through traces and outputs to figure out what actually changed.
This talk covers how clear diffing can make eval results more actionable by integrating into developer’s day-to-day workflows and how to make those diffs dead simple for developers to understand. Drawing from my experience building the eval framework for Sentry’s AI products and many years of building developer tooling, I’ll share practical ways to surface these results directly in developer workflows and make evals something engineers can use in their daily work.read more
8 Ways Production Agents Fail Silently (And the Detection Patterns That Catch Them)
● Applied AI
📍 Theater 4
What if your agent is failing right now, and your monitoring has no idea? Your dashboards are green. Your APIs are healthy. Your latency is normal. The agent returns a perfectly reasonable answer. And yet, something is wrong.What if your agent is failing right now, and your monitoring has no idea?
Your dashboards are green. Your APIs are healthy. Your latency is normal. The agent returns a perfectly reasonable answer.
And yet, something is wrong.
Agentic systems create a new reliability problem: the system can appear healthy while the behavior is quietly degrading. A model can lose important context, take the wrong path, trust bad information, or pass a flawed assumption downstream, all while producing output that looks completely reasonable.
The question is no longer just: Is the system running?
How do you know the agent is doing what you think it is doing?
This talk takes a closer look at the failure patterns that emerge when agents move from demos into real systems, and the engineering discipline needed to catch them before they become expensive incidents.
You’ll take away:
- A practical failure model for recognizing how autonomous systems can go wrong without producing a conventional error.
- A sharper way to evaluate agent behavior when uptime, latency, and error rates tell only part of the story.
- Concrete architectural techniques for containing bad reasoning, validating agent behavior, and preventing one weak step from becoming a larger system failure.
- A reliability mindset for agents: designing for evidence, explicit boundaries, and failures that are visible and recoverable.
Because the hardest failures to debug are the ones that never look like failures.read more
Divya MahajanSoftware Engineer, Amazon
☆
Wed 9/30 · 2:30 PM - 2:55 PM
Act - Verify - Refine: The Future of Adaptive Investment Agents
● AI Strategy
📍 Theater 5
In financial asset management, “good enough” AI introduces significant risk. While Large Language Models (LLMs) can generate market narratives at scale, hallucinations and unverifiable claims limit their use in regulated settings.In financial asset management, “good enough” AI introduces significant risk. While Large Language Models (LLMs) can generate market narratives at scale, hallucinations and unverifiable claims limit their use in regulated settings. To address this, Harbor Capital Advisors develops a Multi-Agent Act - Verify - Refine (AVR) framework - a self-correcting, validation-driven architecture designed for reliable, auditable outputs.
We present how manual investment workflows can be transformed into agentic systems using a closed-loop design. Specialized agents decompose outputs into atomic claims, verify them against structured data sources, and iteratively refine results until they meet predefined standards. We describe the design of validation-centric agent workflows, automated claim verification, and data grounding, along with targeted human-in-the-loop (HITL) interventions that balance expert oversight with scalability.read more
Shuen MeiVP, Dir of Software and Data Engineering, Harbor Capital Advisors
☆
Wed 9/30 · 2:30 PM - 2:55 PM
Evaluating Open Models on Production Traffic
● The AI Stack
📍 Theater 3
Open models now match closed frontier models on quality at a fraction of the cost. Choosing one is the easy part.Open models now match closed frontier models on quality at a fraction of the cost. Choosing one is the easy part. Running it well is where teams get surprised, usually under sustained load, when behavior that looked fine in evaluation stops holding.
Together AI serves these models at production scale, across thousands of GPUs and hundreds of millions of requests. Ollama is the front door for developers who use them every day. This session covers both vantage points: what it takes to serve an open model efficiently at volume, and what developers hit once they start building on one. Time for questions at the end.read more
Zain HasanStaff AI/ML Engineer, DX, Together AI
Parth SareenMember of Technical Staff, Ollama
☆
Wed 9/30 · 2:30 PM - 2:55 PM
Speed Meets Supervision: The Value of Human Judgement in an Agentic World
● AI Frontiers
📍 Theater 1
As enterprise adoption of AI accelerates, organizations may struggle to balance engineering speed with system reliability and human oversight. In this talk, Jennifer Strong and Snap Inc.As enterprise adoption of AI accelerates, organizations may struggle to balance engineering speed with system reliability and human oversight. In this talk, Jennifer Strong and Snap Inc. CIO & SVP of Engineering, Saral Jain, discuss how AI agents can support human employees and explore practical strategies for embedding human intuition, experience, and judgment into every stage of AI implementation.read more
Saral JainSVP of Engineering and Chief Information Officer, Snap Inc.
Jennifer StrongTech Journalist, SHIFT
☆
Wed 9/30 · 2:30 PM - 2:55 PM
The Agentic Shift: How to Build Products When AI Can Actually Do Things
● AI Builders
📍 Theater 2
We've moved from AI that generates answers to AI that takes actions - and it's broken most of our existing product design playbooks.We've moved from AI that generates answers to AI that takes actions - and it's broken most of our existing product design playbooks. This talk shares concrete lessons from building agentic AI products at Google Labs, including Project Mariner (an autonomous web agent), Opal (chained AI workflows), and Pomelli (AI marketing for SMBs), alongside a personal experiment: orchestrating a team of 14 AI agents for everything from meal planning to publishing a coloring book on Amazon KDP.
The hard-earned lesson? Agentic products live in a year of tensions - exciting vs. overwhelming, autonomy vs. control, thinking big while shipping fast - and the teams that ship well are the ones who design for these tensions instead of pretending they don't exist.
What attendees will take away:
• A risk framework for agent autonomy based on reversibility, blast radius, and risk type (reputational, monetary, legal) - because "when should the agent ask permission?" is the central design question and the answer is never one-size-fits-all
• The "Do It With Me" vs. "Do It For Me" spectrum - a practical model for calibrating how much user involvement any given task should have, with real examples from shipping
• Why designing for imperfection matters more than designing for accuracy - agents are non-deterministic systems, and the winning design principle is humans-in-the-loop as a feature, not a fallback
• Iterative co-discovery: how users and agents think together, and why scope expansion through collaboration is not scope creep - it is the product working as intended
• A proactive design quadrant for finding magic (not Clippy) - when proactive agents feel delightful vs. intrusive, based on task value and user effort
This talk is grounded in real product work, real failures, and the honest gap between "AI agent demo" and "AI agent daily driver." No product pitches - just frameworks and lessons from building at both Google scale and personal scale.read more
Jaclyn KonzelmannProduct Dir, Google
Built for you by(617)-PATRICK, Champion AI speaker
Everyone says they're "transforming" with AI. Few can prove it.Everyone says they're "transforming" with AI. Few can prove it. This panel puts two operators who actually have receipts in the room - Claire Southey, Chief AI Officer at Rokt, running AI strategy inside a platform processing billions of real-time transactions, and Todd Earwood, CEO of Betra Works, with a track record of ventures built through this exact shift. Moderated by Jason Hiner (The Deep View), expect real examples, real failures, and zero buzzwords.read more
Claire SoutheyChief AI Officer, Rokt
Jason HinerChief Content Officer and Editor-in-Chief, The Deep View
Todd EarwoodCEO, Betra Works
☆
Wed 9/30 · 3:00 PM - 3:25 PM
Cyber Offense Hits Its Moore’s Law Moment - when compute gets cheap, persistence becomes the exploit.
● AI Strategy
📍 Theater 5
AI is pushing cyber offense into a new regime: when compute and automation become cheap, the attacker’s marginal cost per attempt collapses.AI is pushing cyber offense into a new regime: when compute and automation become cheap, the attacker’s marginal cost per attempt collapses. The result is not merely “better tools,” but a structural change in attacker economics, high-variance, high-volume iteration becomes the advantage, and persistence itself starts to function like an exploit. This session presents that shift using real-world signals, including CVE trends and disclosure patterns, bug-bounty reward dynamics, and observed movement in underground markets (pricing, supply, and demand), to explain what is changing and why the implications are durable.
The talk organizes the transformation into three dominant vectors:
1.Agentic Attack Orchestration - automation at scale, where planning, chaining, and rapid adaptation become commodity capabilities.
2. Autonomous Vulnerability Discovery & Exploitation - accelerated discovery of weak seams across complex, fast-changing systems.
3. AI-Driven Persuasion & Social Engineering - optimized influence at scale across human and organizational interfaces.
A key question explored is what happens to defense when the adversary can run orders of magnitude more experiments per unit time and cost. In that environment, snapshot assurance, periodic reviews, one-off penetration tests, static checklists, becomes less predictive. The core argument of the session is that red teaming only works at this new tempo when it is operationalized as a continuous measurement and control loop: adversarial scenarios are translated into repeatable security evaluations, those evaluations become regressions that run through change (models, prompts, tools, policies, and integrations), and the resulting signals drive release gates, monitoring, and governance priorities.
The session is explicitly not a “how to attack” talk. It stays at the architecture, measurement, and governance layer, focusing on defensive design patterns, evaluation approaches, and operationalization rather than exploit instructions. It also draws a hard boundary around what red teaming cannot buy on its own: it cannot compensate for weak authorization design, systemic data exposure, or supply-chain dependencies without complementary controls such as least-privilege boundaries, provenance, and robust telemetry.
Finally, the talk introduces defensive deception as a cross-cutting capability that connects red teaming to day-to-day detection: honeytokens, canary artifacts, and instrumented decoy surfaces that turn attacker-scale automation into high-signal measurements, improving detection fidelity and accelerating containment across all three vectors.
Attendees will leave with (a) a data-informed model of how attacker economics is shifting, (b) a practical blueprint for running AI red teaming as continuous evaluation tied to engineering and governance decisions, and (c) a clear prioritization lens for defensive investment as the other two vectors mature.read more
Julia DiamentPrincipal AI Architect, Office of the CTO, Microsoft
☆
Wed 9/30 · 3:00 PM - 3:25 PM
Designing Durable Conversation State for AI Agents: Production lessons from building Intuit's A/B testing agent
● AI Builders
📍 Theater 2
At Intuit, we run AI agents that help teams create and manage A/B experiments through natural language.At Intuit, we run AI agents that help teams create and manage A/B experiments through natural language. These agents operate within a multi-agent system, receiving requests from upstream orchestrators, calling external APIs via tool use, and maintaining stateful conversations across multiple turns. Getting the agent to work was the easy part. Getting it to survive production, with automated agents sending rapid-fire requests, external APIs with undocumented constraints, and gateway timeouts killing requests mid-execution, was where the real engineering happened. In this talk, I'll share production patterns we developed from real failures: 1. Field descriptions are your most reliable contract with the LLM. Tool descriptions can be truncated by clients and frameworks. But Pydantic field descriptions become part of the JSON schema. They're never truncated, always reach the LLM, and are used for client-side validation before the call even hits your server. A missing precision constraint on a single field caused our LLM to retry a failing API call 12 times in 48 seconds. Moving the constraint into the field description fixed it in one try. I'll show why "field first" is the rule for any tool, whether it's a LangGraph tool, an MCP server, or a raw OpenAI function call. 2. Error message quality determines whether your agent self-corrects or spirals. In the same incident, the LLM received two different errors. One said "must not contain spaces" and was fixed in 1 try. The other said "invalid value 33.33" with no explanation and was retried 12 times. Same LLM, same tool, same conversation. The only difference was how specific the error was. I'll show how actionable error messages are the cheapest reliability investment you can make. 3. Stateful agents have a hidden failure mode nobody talks about. When a request dies mid-execution (timeout, crash, OOM), your checkpointed state can capture an incomplete conversation that violates the OpenAI tool call contract. Every future request on that thread fails permanently with no recovery path. I'll show exactly how this happens and how to design for self-healing state. 4. Intent classification needs a safety net. We use an LLM to classify user intent and dynamically bind only the relevant tools per turn, reducing the chance of the LLM calling the wrong tool entirely. But LLM classification isn't 100% reliable, so we layer heuristic fallbacks underneath. I'll share the dual-layer routing pattern and why limiting tool visibility per turn prevents entire categories of errors. Each pattern comes from a real production incident. I'll walk through the failures, the wrong assumptions we made during debugging, and the concrete code changes that fixed them. Key takeaways: -Put constraints in Pydantic field descriptions, not just tool docstrings. Field descriptions are the only channel guaranteed to reach the LLM untruncated. -Invest in actionable error messages. They're the difference between 1-try self-correction and infinite retry loops. -Design for mid-execution failures in any stateful agent system. Timeouts will corrupt your conversation state. -Layer your intent routing. LLM classification plus heuristic fallbacks. Target audience: Software engineers, ML engineers, and tech leads building or operating AI agents with tool use in production. The patterns apply regardless of framework (LangGraph, Google ADK, CrewAI, MCP, raw function calling). Evidence of expertise: I've been selected to speak at the MLOps Community AI Agents Summit (April 2026) on this topic. I've also published on the MCP request lifecycle and tool schema design: https://medium.com/@vidyameenakshi/the-mcp-request-lifecycle-what-actually-happens-when-an-ai-agent-calls-your-tool-57bddc2211ca Adaptability: I'm happy to receive feedback and adjust the talk to best fit the conference's themes and audience. The content can be tailored for a 20-minute deep dive or a 30-minute session with extended Q&read more
Vidya MeenakshiSenior Software Engineer, Intuit
☆
Wed 9/30 · 3:00 PM - 3:25 PM
Lyrics: The Language Model of Music
● Applied AI
📍 Theater 4
At Musixmatch - the lyrics database powering Spotify, Apple Music, and most of the streaming ecosystem - we recently shipped lens.musixmatch.com, an AI agent that lets labels and publishers query their catalog as if it were a knowledge base.At Musixmatch - the lyrics database powering Spotify, Apple Music, and most of the streaming ecosystem - we recently shipped lens.musixmatch.com, an AI agent that lets labels and publishers query their catalog as if it were a knowledge base. This talk is about how we built it, what broke, and what the music catalog problem reveals about building production AI on rights-managed, culturally dense corpora.
The core insight is this: lyrics are music's native language model. For over a century, the music industry has been generating structured semantic data - mood, theme, narrative, entity, temporal context, emotional arc - embedded in song text. Yet metadata extraction in music has historically been restricted to the audio signal: tempo, key, genre, acoustic mood. The lyrical component has been almost entirely ignored as a retrieval and classification surface. That's the gap we closed.
We built a system with four functional layers.
The first is a metadata extraction pipeline that uses LLMs and custom prompt engineering to pull structured information from raw lyrics at scale: themes, ratings, named entities (brands, places, people, events), emotional tone, and cultural references - with entity extraction that understands context rather than relying on string matching, so "Ford" in a country song resolves correctly whether it refers to a person, a truck, or a river crossing.
The second is a hybrid search index built on OpenSearch, combining dense vector embeddings with structured metadata filters, which allows queries that neither pure semantic search nor keyword search can handle alone - "find songs about corporate greed appropriate for a retail brand campaign, released after 2018" requires both.
The third layer is where it gets architecturally interesting: a natural language querying system built on an agentic framework we call Domino, where user intent expressed in plain language is decomposed by an LLM into a structured DSQL query via tool calls and in-context learning. The agent doesn't just retrieve - it explains its reasoning, handles ambiguity, and chains multiple retrieval steps when a single query is insufficient. We'll walk through specific failure modes: what happens when cultural context is underspecified, when a user's brief is contradictory, and when the model's knowledge of a cultural reference predates the song that contains it.
The fourth layer is trend and classification analysis: using the extracted metadata to map an artist's thematic evolution across albums, identify cultural moments in a catalog (which songs align with an upcoming World Cup, a brand's new campaign, a breaking news cycle), and surface royalty and DSP performance signals alongside content intelligence in a single interface.
The hardest problems were not model selection. They were: defining what "derived data" means as a technical specification when the legal definition didn't previously exist; building evaluation pipelines for tasks where ground truth is cultural and subjective; and designing a retrieval architecture where no raw lyric content ever surfaces in output - a hard licensing constraint that shaped every architectural decision downstream.
Attendees will leave with concrete patterns for building hybrid semantic search on specialized corpora; a practical framework for agentic NLQ systems using tool calls and ICL; honest benchmarks on where general-purpose LLMs degrade on culturally specific data; and a replicable approach to building AI agents on licensed, rights-restricted datasets - a problem that music faces first, but that every regulated domain will face next.read more
Marco PagliaPresident - Product & Engineering, Musixmatch
☆
Wed 9/30 · 3:00 PM - 3:25 PM
The Content Machine: Orchestrating Content Workflows at Netflix Scale
● The AI Stack
📍 Theater 3
Every title that reaches Netflix's millions of subscribers passes through a gauntlet of server-side decisions - and with Netflix's catalog growing faster than ever before, the backend systems powering those decisions have never been under more pressure.Every title that reaches Netflix's millions of subscribers passes through a gauntlet of server-side decisions - and with Netflix's catalog growing faster than ever before, the backend systems powering those decisions have never been under more pressure.
Most teams solve this by stitching together separate systems for automation, ML, and human review. It works - until it doesn't. At Netflix, we took a different approach: we built a single backend framework where all three are first-class citizens in the same pipeline, sharing the same API contracts, the same failure model, and the same observability story.
In this talk, we'll get into the real battle-tested engineering behind it - how we designed microservices that stay stable when components swap out, how we use in-memory data architectures to sidestep the performance ceiling of traditional distributed databases, and how we keep API contracts clean across heterogeneous processing stages. We'll also show how this foundation naturally extends to AI agents - not as a future vision, but as a direct consequence of the architectural decisions we made today.
Expect concrete patterns, hard-won lessons, and a few surprises from running this in production at Netflix scale.read more
Beyond Hyperscale: The New Era of Enterprise Infrastructure
● AI Strategy
📍 Theater 5
Every application ever built needs to be rebuilt - with agentic AI at the core. Moving from cloud-native to ai-native means a new architecture for unlocking enterprise data to safely, securely deliver personalized experiences and services at scale.Every application ever built needs to be rebuilt - with agentic AI at the core.
Moving from cloud-native to ai-native means a new architecture for unlocking enterprise data to safely, securely deliver personalized experiences and services at scale.
And delivering business outcomes requires delivering those experiences to employees and customers globally, wherever they live and work.
To build the Agentic Enterprise, enterprises need to rethink their infrastructure strategy, own their data, and own their AI stack.
In this session, we’ll explore a new infrastructure approach to unlock enterprise AI - and how agents can help.read more
Kevin CochraneCMO, Vultr
☆
Wed 9/30 · 4:00 PM - 4:25 PM
Bobium Brawlers: Shipping the First Mobile Game with Fully On-Device Generative AI
● The AI Stack
📍 Theater 3
Most AI applications today rely on cloud inference, which introduces latency, privacy concerns, operational cost, and limits real-time interactivity. In games, these constraints are especially problematic: player experiences must respond instantly and scale to millions of users without bankrupting developers.Most AI applications today rely on cloud inference, which introduces latency, privacy concerns, operational cost, and limits real-time interactivity. In games, these constraints are especially problematic: player experiences must respond instantly and scale to millions of users without bankrupting developers.
In this talk, we present the architecture and production lessons behind Bobium Brawlers, the first mobile game that runs generative AI entirely on device. The system powers real-time creature and gameplay content generation without a deluge of cloud API calls.
We will walk through the design of a production system capable of running generative models locally on consumer devices, including:
Designing gameplay systems around local LLM and generative model inference
Techniques for model compression, optimization, and quality
Integrating AI generation directly into core gameplay loops rather than cosmetic features
Handling latency, guardrails and reliability when inference happens inside the game
Lessons learned from deploying the system to real players and observing emergent gameplay
We will also discuss how AI-native design changes traditional game development workflows, including content pipelines, player creativity systems, and balancing mechanics when AI becomes part of the game itself.
Attendees will walk away with practical insights into shipping AI-first interactive systems on constrained devices, including architectural patterns, optimization strategies, and lessons from deploying generative AI in a live product.read more
Piero MolinoCEO, Studio Atelico
☆
Wed 9/30 · 4:00 PM - 4:25 PM
Every Agent, Everywhere, All at Once
● Applied AI
📍 Theater 4
Coding agents are deaf to anything outside their own session, and a LangGraph or CrewAI one has no idea the others exist. Different vendors, different frameworks, different machines none of them share a way to work together.Coding agents are deaf to anything outside their own session, and a LangGraph or CrewAI one has no idea the others exist. Different vendors, different frameworks, different machines none of them share a way to work together. This demo fixes that live: the Claude Code on your laptop, Codex on your colleague's, a LangGraph agent you're running locally, and the OpenClaw on your Mac Studio at home collaborating on the same goal, going back and forth, full-duplex, across every vendor, framework, and machine line at once.read more
Vlad LuzinCTO & Co-founder, BAND
☆
Wed 9/30 · 4:00 PM - 4:55 PM
Startup Showdown
● Events & Social
📍 Theater 1
Eight emerging AI startups take the main stage for fast-paced, four-minute pitches in front of active investors and a live audience. With real-time judge scoring and audience voting, founders compete for Top Startup and the People’s Choice Award.
Brady ForrestFounder, Ignite Talks
Emilio AndereCo-Founder/CEO, Wafer
Matan Bar-EfratCo-founder and CEO, Rein Security
Maya GosztylaCo-Founder and Chief Scientific Officer, BrainStorm Therapeutics
Akshat JaggaFounder & CEO, Datapoint AI
Justin PoehneltFounder and CTO, Helen’s Foundry
Anthony ScodaryCo-founder, Guava
Sumanyu SharmaFounder & CEO, Hamming AI
Zach DeWittPartner, Wing VC
Max AbramsPartner, Scale
Nick GiomettiSenior Principal, B Capital
☆
Wed 9/30 · 4:00 PM - 4:25 PM
Why Your AI Keeps Getting It Wrong: The Semantic Backbone Fix
● AI Builders
📍 Theater 2
Most enterprise AI failures aren't model failures: they're meaning failures. When LLMs don't share your business's definitions of "customer," "risk," or "compliant," every downstream decision inherits that ambiguity.Most enterprise AI failures aren't model failures: they're meaning failures. When LLMs don't share your business's definitions of "customer," "risk," or "compliant," every downstream decision inherits that ambiguity. This session introduces the Semantic Backbone: a knowledge infrastructure that enforces a single, governed semantic agreement across fragmented data silos, grounding LLMs in the business logic that actually matters.
We'll show how Semantic GraphRAG architectures reduce hallucinations by up to 90% and make AI reasoning explainable and auditable, without rebuilding your data stack. Data and AI leaders will leave with a concrete framework for moving from brittle prototypes to production-grade, domain-aware AI systems.
- Diagnose why fragmented data definitions are the root cause of most enterprise AI failures in production
- Learn how an enterprise knowledge graph and a context graph play together
- Explore the Semantic Backbone as a governed knowledge infrastructure that enforces unified business logic and context across silos, giving LLMs a reliable foundation to reason from
- Examine how Semantic GraphRAG delivers explainable, auditable AI outputs, and what a practical transition from prototype to production looks like for data leadersread more
Rick CraneSr VP of Revenue (CRO), Graphwise
Built for you by(617)-PATRICK, Champion AI speaker
Deploying LLMs in Production to enhance grocery shopping
● Applied AI
📍 Theater 4
This session explores the technical challenges and design patterns behind deploying large language models (LLMs) in real‑world, high‑scale consumer applications. The talk will cover two distinct but related systems: an AI‑driven recipe generation pipeline and a production‑grade LLM‑powered support assistant.This session explores the technical challenges and design patterns behind deploying large language models (LLMs) in real‑world, high‑scale consumer applications. The talk will cover two distinct but related systems: an AI‑driven recipe generation pipeline and a production‑grade LLM‑powered support assistant.
The recipe generation workflow examines how SEO‑driven keyword inputs are transformed into structured, high‑quality recipes through multi‑stage LLM prompting, retrieval‑based enrichment, and validation steps. It also discusses how image‑generation models are integrated into the pipeline and how quality and consistency are maintained across large content batches.
The support assistant portion focuses on the architecture and reliability requirements of a large‑scale LLM chatbot serving millions of queries. Topics include context management, retrieval‑augmented responses, multi‑tenant considerations, metadata scoping, error surfaces unique to model providers, and the tracing and observability techniques required to identify and remediate production failures.
Attendees will gain a practical understanding of how LLM systems are built and maintained at scale, how semantic failures differ from traditional software failures, and what architectural patterns can improve accuracy, predictability, and operational reliability in production AI applications.read more
Yogiraj AwatiEngineering Manager, Instacart
☆
Wed 9/30 · 4:30 PM - 4:55 PM
Lessons from Building a $100M Product in Six Weeks at OpenAI
● AI Builders
📍 Theater 2
Engineering is shifting from synchronous, line-by-line implementation to an asynchronous model in which agents execute, verify, and retry while humans own intent, architecture, constraints, and release decisions.Engineering is shifting from synchronous, line-by-line implementation to an asynchronous model in which agents execute, verify, and retry while humans own intent, architecture, constraints, and release decisions.
Drawing on my experience building OpenAI Ads through Codex from day one and scaling it to more than $100 million in ARR in under six weeks. I’ll show why the advantage does not come from a magic prompt. It comes from the system around the model: persistent workspaces, connections to real engineering tools, reusable skills, objective verification loops, and clear human boundaries around production.
Attendees will leave with a practical framework for moving faster with coding agents without surrendering engineering judgment or accountability.read more
Brian YangMember of Technical Staff, OpenAI
☆
Wed 9/30 · 4:30 PM - 4:55 PM
Measuring What Matters: Statistical Rigor in AI Accuracy Evaluations
● The AI Stack
📍 Theater 3
Most teams evaluate AI accuracy by running a handful of test queries, picking whichever system looks best, and shipping.Most teams evaluate AI accuracy by running a handful of test queries, picking whichever system looks best, and shipping. The problem is that ad-hoc testing doesn't reflect production behavior and isn't replicable; worse, standard benchmarks aren't calibrated to specific use cases. When you're making infrastructure decisions that cost hundreds of thousands of dollars, cherry-picked demos are not a measurement strategy.
This talk introduces a rigorous framework for evaluating AI search quality that applies classical ML statistical principles to a problem the industry has largely treated as subjective. I’ll cover how to construct a golden test set that actually predicts production performance, how to validate LLMs-as-judge against human raters before trusting them at scale, and how to run controlled comparisons across providers.
The deeper argument is about what most evaluations get wrong at the measurement level. LLM-based evaluation is inherently stochastic, so the same query, run through the same system, produces different outputs across trials. Reporting accuracy alone can't tell you whether a performance difference is real or just noise; I’ll explain how to decompose evaluation variance into signal and instability. You’ll leave with a methodology you can apply immediately to your own vendor comparisons and internal eval pipelines.read more
Zairah MustahsanApplied Scientist, You.com
☆
Wed 9/30 · 4:30 PM - 4:55 PM
One of the Most Complex Thing Humans Build: How AI Is Finally Cracking Chip Design
● AI Strategy
📍 Theater 5
Every smartphone, AI accelerator, and data center processor starts as a design problem so complex it defies human intuition.Every smartphone, AI accelerator, and data center processor starts as a design problem so complex it defies human intuition. A modern chip contains billions of transistors packed into a space smaller than a fingernail, designed to operate at speeds measured in trillionths of a second, across dozens of operating conditions simultaneously and it has to be right the first time. A single undetected error means scrapping wafers that cost millions of dollars and losing six to twelve months before a corrected chip can be manufactured. There are no patches, no hotfixes, no rollbacks. You tape out, or you do not ship.
This is the environment where AI is now being deployed and where the gap between what AI promises in a demo and what it delivers under real pressure is most brutally exposed.
This session examines what it actually takes to move AI from prototype to production in one of the most demanding engineering disciplines on earth, built on production experience designing AI and AR chips at advanced process nodes.
The first question is where AI delivers genuine value in chip design and where it does not. AI can predict timing problems before a layout is finalized, flag high-risk areas before they become emergencies, and explore design tradeoffs at a scale no human team could match manually. In the right places, production data shows 15 to 30 percent gains in closure time. But AI also fails quietly in ways more dangerous than obvious failures - and knowing where the boundary sits is as important as knowing the capability exists.
The second question is what production-ready actually means. In chip design, a tool that requires a new interface or workflow change will not get used under tapeout pressure. Production-ready means invisible: one configuration flag in an existing flow, graceful recovery when compute infrastructure fails, and runtime predictability engineers can rely on when schedules are tight. These are not quality-of-life improvements - they are the entry criteria for adoption.
The third question is what it costs to keep these systems working after deployment. AI systems in chip design do not stay current on their own. Chip physics change with each new process node. Design tools update and break data pipelines in ways that surface at 2 a.m. during critical runs. Models drift as design styles evolve. Sustaining AI in this environment requires dedicated ownership and organizational alignment - and teams that do not plan for this find their successful prototypes quietly stop working within two or three product cycles.
The fourth question is how senior practitioners actually evaluate these tools - which is very different from how they perform in a demo. They are not looking for the best benchmark score. They are asking whether they can debug it at midnight, whether it holds up on the hardest late-stage blocks, and whether it fails gracefully or catastrophically when inputs are imperfect.
The fifth question is what success looks like at scale. When AI works in chip design, senior engineers reach for it by default on the hardest problems, schedules become more predictable, and silicon quality improves without a proportional headcount increase. Organizations that get there treat sustainment and cross-team alignment as first-class investments - not afterthoughts.
Attendees will leave with a framework for evaluating where AI is worth the complexity in their own domain, what commitments separate deployments that scale from those that stall, and how to lead a rigorous conversation about AI ROI that goes beyond slide-deck metrics.read more
☆
Wed 9/30 · 4:30 PM - 4:55 PM
When will biology have its ChatGPT moment?
● AI Strategy
📍 Theater 5
In life sciences, AI models haven't had an internet's worth of biological data to learn from. We review what's still missing - and what it will take for biology to have its ChatGPT moment.
Party in the Park is an open-air networking celebration bringing the AI community together for an evening of live music, food and drinks, games, and a special surprise announcement. All registered attendees are invited to attend.Party in the Park is an open-air networking celebration bringing the AI community together for an evening of live music, food and drinks, games, and a special surprise announcement. All registered attendees are invited to attend.
Sponsored by Fireworks and Merge.read more
Ignite Talks is a fast-paced evening of short, high-energy presentations designed to spark new ideas, unexpected conversations, and fresh perspectives.Ignite Talks is a fast-paced evening of short, high-energy presentations designed to spark new ideas, unexpected conversations, and fresh perspectives. Join fellow attendees for a fun and engaging close to Day 1, featuring a lineup of rapid-fire talks from across the AI community.read more
THURSDAY, OCTOBER 1 (DAY 2)60 SESSIONS
Built for you by(617)-PATRICK, Champion AI speaker
The Post-Transformer Frontier: More Intelligence per Watt, Token, and Dollar
● Keynotes
Keynote
📍 Theater 1
Transformers have reshaped AI, but they were never designed as a foundation for intelligence that must persist and grow over time. Today’s reasoning models often gain capability by generating and accumulating an ever-growing trace of reasoning tokens.Transformers have reshaped AI, but they were never designed as a foundation for intelligence that must persist and grow over time. Today’s reasoning models often gain capability by generating and accumulating an ever-growing trace of reasoning tokens. This raises latency, drives up inference cost, and increases energy consumption.
We are already seeing diminishing returns from scaling parameters and inference-time compute. This keynote explains why the next frontier in AI will come from architectural breakthroughs that produce more intelligence per watt, token, and dollar.
I will map the emerging post-Transformer landscape across three architectural frontiers: persistent and recurrent memory, sparse and selective computation, and continual adaptation - and show how they need to come together to achieve this goal.
I will then introduce BDH - an architecture that unifies memory, reasoning, and adaptation - and lay out a path beyond static parameters and ephemeral token histories toward AI systems that retain what matters, reason and adapt over long horizons, and improve through experience.read more
Adrian KosowskiChief Scientific Officer and Co-founder, Pathway
☆
Thu 10/1 · 9:25 AM - 9:40 AM
Your Agent's Next Thousand Tools
● Keynotes
Keynote
📍 Theater 1
Every agent framework gives you a loop. What actually matters is what's on the other end of the tool call.Every agent framework gives you a loop. What actually matters is what's on the other end of the tool call. In this keynote, Mike Chambers traces the line from software agents to physical ones, same code, same patterns, and shows why the developers who build the next thousand tools will matter more than those who build the next agent framework.read more
Mike ChambersSenior AI specialist DA, Amazon Web Services
Foundation models have transformed how we interact with AI systems. Can the same transformation take place in robotics, providing us with robots that can be tasked to perform any job through prompts and instructions?Foundation models have transformed how we interact with AI systems. Can the same transformation take place in robotics, providing us with robots that can be tasked to perform any job through prompts and instructions? Getting the same kind of generality and interactivity out of robots that we've come to expect from AI assistants and generative AI tools will require both technical innovations and a sophisticated industrial effort. In this presentation, I will discuss what robot foundation models are, how they have advanced in recent years, and the technical directions that will lead to the most significant advances in the near future.read more
Sergey LevineAssociate Professor, University of California, Berkeley
☆
Thu 10/1 · 9:55 AM - 10:05 AM
Shortlisting and Evaluating Open Models
● Keynotes
Keynote
📍 Theater 1
There are dozens of open models worth considering and no reliable way to know which ones suit your workload. General rankings will not tell you, and testing all of them is not realistic.There are dozens of open models worth considering and no reliable way to know which ones suit your workload. General rankings will not tell you, and testing all of them is not realistic. So teams pick a few, run an evaluation, and hope the result means something.
This session covers how to narrow the field and how to build an evaluation that predicts production behavior. What representative workload data looks like. Which performance characteristics matter before a model goes live. Rochelle works on these migrations across Together AI's customers and will go through what the successful ones did differently.read more
Rochelle MatternDir, Field Engineering, Together AI
As adoption of AI agents increases, more businesses and individuals are realizing the risk involved in exposing IP, proprietary and financial data, and personal information to centralized AI providers.As adoption of AI agents increases, more businesses and individuals are realizing the risk involved in exposing IP, proprietary and financial data, and personal information to centralized AI providers. Illia Polosukhin, co-author of “Attention Is All You Need,” calls for a shift towards “user-owned AI,” or AI systems that optimize for users’ agency and preserve their privacy. He unpacks his vision for an open AI stack and the requirements for more performant and trustworthy AI systems with verifiable compute, globally available to all.read more
Illia PolosukhinCo-Founder, NEAR AI
☆
Thu 10/1 · 10:20 AM - 10:30 AM
Specialized Intelligence with Open Models
● Keynotes
Keynote
📍 Theater 1
Rob Ferguson will discuss how organizations figure out their real AI trade-offs: when to post-train, when to keep paying for frontier models, and how to know the difference.Rob Ferguson will discuss how organizations figure out their real AI trade-offs: when to post-train, when to keep paying for frontier models, and how to know the difference. The answer comes from eight practices that turn AI from a line item into a competency.read more
Reality Is the Final Verifier: On Two Key Gaps in Agentic Software Engineering
● Keynotes
Keynote
📍 Theater 1
AI coding agents can generate and test software orders of magnitude faster than humans can. Yet they optimize against requirements and environment models that only approximate stakeholder intent and the real world.AI coding agents can generate and test software orders of magnitude faster than humans can. Yet they optimize against requirements and environment models that only approximate stakeholder intent and the real world. This talk introduces the two-gap framework: the requirement gap and the model gap. Neither can generally be closed in an open, changing world. Although these gaps are longstanding, agents make them far more consequential because they may lack relevant context, iterate much faster than humans, and automatically test and deploy implementations. Reliable agentic development therefore requires an outer assurance - revision loop that uses deployment outcomes and human judgment to continually narrow the gaps.read more
Ion StoicaBerkeley Electrical Engineering & Computer Sciences, Professor & Co-founder, UC Berkeley, Arena, Infarct, and SkyPilot
Built for you by(617)-PATRICK, Champion AI speaker
Behind the Endpoint: Running Production Inference at Scale
● The AI Stack
📍 Theater 3
Today’s most ambitious AI products combine multiple models, custom and post-trained models, multimodal workloads, and long-running agents - often under unpredictable demand. At the same time, users still expect them to feel instantaneous, reliable, and economically sustainable.Today’s most ambitious AI products combine multiple models, custom and post-trained models, multimodal workloads, and long-running agents - often under unpredictable demand. At the same time, users still expect them to feel instantaneous, reliable, and economically sustainable.
Drawing on patterns across AI-native startups and large enterprises, this session will unpack what it takes to run inference well at production scale: matching models to the right engines and hardware, scaling without painful cold starts, managing latency, reliability, and cost tradeoffs, and continuously adapting as models and workloads evolve.
More importantly, we will examine how these infrastructure decisions shape the customer experience, and how AI architectures are evolving behind the simplicity of a model endpoint. As inference becomes more central to product differentiation, companies are beginning to treat it not as backend infrastructure, but as a core product capability.read more
Sai MaddaliProduct Lead, Baseten
☆
Thu 10/1 · 11:30 AM - 11:55 AM
From 200B Daily Events to Autonomous Agents: It Was Never About the Model
● Applied AI
📍 Theater 4
The hardest part of shipping autonomous agents isn't the agent framework - it's everything underneath it. In six months, AppsFlyer went from AI experimentation to a production Agent Platform with 600+ customers running thousands of autonomous marketing agents every week.The hardest part of shipping autonomous agents isn't the agent framework - it's everything underneath it.
In six months, AppsFlyer went from AI experimentation to a production Agent Platform with 600+ customers running thousands of autonomous marketing agents every week. These agents analyze campaign performance, surface optimization opportunities, and act on recommendations - autonomously, at enterprise scale, across some of the world's largest advertisers. This talk is about what made that velocity possible - and it wasn't the model.
We'll walk through how we built a Google Cloud-native AI platform to transform 200B+ raw daily events into governed, discoverable data products. The critical design decision: an intelligent metadata layer that gives agents the semantic context to reason over enterprise data without hallucinating - turning "data lake" into "data products agents can actually use."
You'll see the specific architectural patterns we used, the governance and design tradeoffs we made to ship fast without losing trust, and the failure modes we hit when agents met messy real-world data. We'll share the patterns that generalize beyond marketing to any domain where agents need to operate safely on complex enterprise data.
This is a practitioner talk. No slides about what agents could do someday. Just the architecture, decisions, and lessons from scaling autonomous systems in production.read more
Inna WeinerVP Product, AppsFlyer
☆
Thu 10/1 · 11:30 AM - 11:55 AM
Fundraising & the Future of AI Companies
● AI Frontiers
📍 Theater 1
In this panel we'll explore where AI capital is actually going, and what investors are underwriting when the underlying technology changes every quarter.In this panel we'll explore where AI capital is actually going, and what investors are underwriting when the underlying technology changes every quarter. We'll dig into how valuations are being justified at the earliest stages, what founders are being asked to prove before a term sheet, where defensibility really comes from when models are a commodity, and what the next 18 months look like for companies raising into this market.read more
Lauri MoorePartner, Bessemer Venture Partners
Jon SakodaFounding Partner, Decibel
Matthew StrozierBureau Chief, WSJ Pro, The Wall Street Journal
☆
Thu 10/1 · 11:30 AM - 11:55 AM
When Agents Meet Creativity: What Works and What Breaks
● AI Builders
📍 Theater 2
Creativity is a harder domain for agents than task execution because it tests taste, ambiguity, judgment and trust. This talk shares what works, what breaks and what it takes to build agents that genuinely expand human creativity.
Ran LiTechnical Lead, Google
☆
Thu 10/1 · 11:30 AM - 11:55 AM
Why Child Safety Breaks Standard AI Safety, and How We Can Fix it
● AI Strategy
📍 Theater 5
Most AI safety operates under a set of strong assumptions. You can collect examples of harmful content to test your classifiers. You can red-team your model to find failure modes. You can hire human reviewers.Most AI safety operates under a set of strong assumptions. You can collect examples of harmful content to test your classifiers. You can red-team your model to find failure modes. You can hire human reviewers. Concepts that are safe in isolation will be safe when combined. The overall result is what matters; the cost of a single mistake is not that high.
Child safety breaks those assumptions.
Child sexual abuse material is illegal to generate in the US, even for safety testing and research. Standardized evaluation datasets don't exist. Those looking to red-team systems or review user reports face documented psychological harm from exposure.
In this talk, Michael Simpson, Machine Learning Engineer at Thorn, presents findings from Child Safety Necessitates New Approaches to AI Safety, a collaborative research paper between Thorn and Carnegie Mellon University, outlining 15 open problems across the AI development lifecycle that pose tangible safety risks to the current generation of models.
The talk is structured around a provocation for AI builders: how can we build systems that provide robust safeguards against extreme harm? The challenges posed here are early signals of a broader class of problems that the field has yet to solve.
The session covers:
- Why the technical, legal, and ethical constraints of child safety make standard AI safety approaches fail, and what that reveals about the assumptions baked into current safety research
- What gaps exist in the current generation of safeguards and best practices
- Concrete open problems that researchers and AI builders can engage with right now, from generation-free model auditing to robust watermarks to securing AI agents against misuse
- A call to action for the builders in the room: what responsible AI development looks like when the content category is inherently harmful
Attendees will learn:
- Why child safety is a stress test for AI safety methodology, not just a policy problem
- A map of 15 specific, real-world open problems across the AI development lifecycle
- The legal, ethical, technical, and wellness constraints that require new ways of thinking about safety
- Specific areas where AI researchers and builders can contribute, from mechanistic interpretability to model hashing to provenance techniquesread more
Beyond the Developer: Architecting the Agentic Commerce Stack of Tomorrow
● AI Strategy
📍 Theater 5
The first wave of agentic commerce has arrived and it's almost exclusively built for developers.The first wave of agentic commerce has arrived and it's almost exclusively built for developers. Autonomous purchasing scripts, AI procurement bots, agent-to-agent negotiation frameworks: today's tooling assumes a technically sophisticated user who can configure APIs, manage credentials, and navigate raw infrastructure. While appropriate for today’s agentic landscape, eCommerce as a whole will need to adapt and evolve for the world of tomorrow.
Just as developers want to delegate the burden of API key configuration, pay-per-use requests, and more to an AI agent today, the shopper of tomorrow will delegate everyday purchasing decisions to AI agents using simple English commands. This shift demands more than better UX, it requires a protocol stack designed for trust, discoverability, consent, and seamless payment at machine speed.
This talk walks the full agentic shopping lifecycle through five converging standards and shows them working together in a live proof-of-concept that settles real value on Hedera testnet: Agent2Agent Protocol (A2A) for agent-to-agent discovery and negotiation; Agent Payments Protocol's (AP2) signed mandates for capturing verifiable purchase authorization; the Legal Context Protocol (LCP) for making terms, governing law, and dispute recourse discoverable and verifiable; Agent Name Service (ANS) for agent identity, so shoppers and merchants know who they're actually dealing with; and x402 over Hedera for programmable stablecoin settlement with enforced spending caps and guardrails on autonomous spend.read more
Building Agentic AI at PayPal: Platform, Governance, and Agent Design
● AI Builders
📍 Theater 2
An agent that can act on production systems is a new actor in the enterprise. It has credentials, and it can accomplish complex tasks across systems autonomously.An agent that can act on production systems is a new actor in the enterprise. It has credentials, and it can accomplish complex tasks across systems autonomously. That changes what the platform underneath has to provide: identity and authorization for non-human actors, transaction-grade reliability, and observability that treats every agent action as a first-class audited event.
At PayPal we run a fleet of specialized agents on shared infrastructure - agents that write code, deploy models, and monitor agents. They sit on a common substrate: a certified agent/mcp registry, agent-to-agent delegation, unified model routing under latency and cost pressure, enterprise memory that persists across sessions, and a governed skills layer so capabilities are reused rather than rebuilt. We're now extending that substrate across cloud project boundaries, so agents owned by different teams and deployed in different environments can find and delegate to each other. Once agents cross those boundaries, identity and audit become essential. In regulated financial infrastructure, governance, security, and Responsible AI aren't a layer bolted on, they're the substrate the swarm of agents runs on.
This talk walks through that architecture, then gets into the parts that were hard to learn: graph-based orchestration with typed state contracts, how we scaled the agent orchestrator, normalizing two agent protocols into one response contract, human-in-the-loop at the right moments, and agent memory that survives across sessions.
You'll leave with patterns you can apply to your own systems and a clearer view of the platform capabilities these systems rely on.read more
Data Architectures in the Age of AI: Rebuilding the data stack from first principals for autonomous agents
● The AI Stack
📍 Theater 3
Every major computing shift has reshaped the software stack, but one assumption held across all of them: humans were the primary consumers of software and software-generated artifacts. AI agents break this assumption.Every major computing shift has reshaped the software stack, but one assumption held across all of them: humans were the primary consumers of software and software-generated artifacts.
AI agents break this assumption. An autonomous agent determines at runtime what data it needs, interprets it, and takes actions, with no human in the loop to catch misinterpretations.
In this new era, bolting AI features onto existing app-centric, or warehouse- and lakehouse-centric stacks just isn't enough to meet the requirements of these agents. The data layer must be rethought from first principles to enable agents to run critical decision-making workflows using structured operational data, reliably and with confidence.
In this talk, I’ll discuss three imperative changes - automatically generated and maintained business semantics, federation as the default retrieval path, and governance - and show how our new agent-native data platform incorporates these principles.read more
Deepti SrivastavaFounder/CEO, Snow Leopard AI
☆
Thu 10/1 · 12:00 PM - 12:25 PM
Experience Is Abundant. Knowledge Is Distilled.
● AI Frontiers
📍 Theater 1
Generative AI gave us an abundance of content. Agentic AI is giving us something potentially more valuable, while also much more overwhelming: experience.Generative AI gave us an abundance of content. Agentic AI is giving us something potentially more valuable, while also much more overwhelming: experience. Every tool call, coding session, workflow, decision, observation, and robot trajectory leaves behind another trace of what happened, what worked, and what didn't.
But accumulating raw piles of experience isn't the same as learning from it.
This talk explores an emerging layer in AI systems devoted not simply to storing context, but to distilling experience into durable, reusable knowledge. We'll connect ideas appearing across agent memory, semantic layers, knowledge bases, coding agents, context graphs, and even physical AI: resolving repeated observations into entities, discovering relationships and patterns, turning trajectories into reusable skills, and feeding what we've learned back into future action.
A number of knowledge primitives are competing for this layer, from the humble Markdown document and hypertext to vector embeddings, model weights, and knowledge graphs. Graphs, in particular, keep popping up. They may occupy a sweet spot: knowledge that is structured yet flexible, explicit and inspectable, portable across models and agent harnesses, and capable of generating other representations without surrendering the underlying structure. We'll pressure-test that claim, asking when the graph is useful, when it isn't, and whether increasingly capable models might eventually swallow the whole thing.
The last wave of AI was about generation. The next frontier may be about accumulation: building systems that don't just produce more experience, but get better at distilling what it means.read more
Jeremy AdamsSenior Developer Advocate, Neo4j
☆
Thu 10/1 · 12:00 PM - 12:25 PM
The Agentic Product Loop: Why shipping faster has raised the stakes on everything else you do
● Applied AI
📍 Theater 4
Every product team is under pressure to "do AI," but most of its benefits to date have centered around product execution, leaving the rest of the product development lifecycle a green field for the next wave of innovation, and increasingly,...Every product team is under pressure to "do AI," but most of its benefits to date have centered around product execution, leaving the rest of the product development lifecycle a green field for the next wave of innovation, and increasingly, a strategic imperative. Drawing from the front lines of building Productboard Spark, Hubert Palan makes the case that these other phases have stayed unsolved for a specific reason. Beyond the orderly structure of a repo, agents require a shared, interconnected substrate. Hubert explores the architecture that closes that gap, and what separates the average DIY product agent from one you can bet your roadmap on.
You'll learn:
- How to tell real decision-quality gains from superficial AI productivity
- The components of a dependable agentic product system, and why each one is easy to underestimate
- Why "build it ourselves" is harder than it looks, and how to frame build-vs-buy for AI
- Why loop speed, not shipping speed, is the real competitive advantageread more
Hubert PalanCo-Founder/CEO, Productboard
Built for you by(617)-PATRICK, Champion AI speaker
AI and Managed Services: Future Models of Symbiosis with AI
● Lunch Sessions
Breakout Session
📍 Theater 5
AI is transforming managed services from a labor-driven model to an intelligence-driven one.AI is transforming managed services from a labor-driven model to an intelligence-driven one. As organizations embed AI into service delivery, the focus is shifting toward defining clear roles for human expertise, autonomous systems, and business stakeholders while maintaining accountability and operational excellence. This session will explore emerging frameworks for AI governance, performance measurement, and outcome-based service delivery, with a particular focus on maximizing ROI from AI investments. Attendees will learn how leading enterprises are optimizing model selection, controlling token costs, and creating sustainable operating models that balance automation, trust, and business value. The session will provide a forward-looking view of how AI and managed services will evolve together to create more efficient, resilient, and scalable enterprises.read more
Dr. Kanav KaholChief AI Officer, Kagen.ai
☆
Thu 10/1 · 12:30 PM - 12:55 PM
Performance Engineering in the Age of AI
● Lunch Sessions
Mainstage Talk
📍 Theater 1
Profiling has long been an esoteric skill that only few people ever took the time to learn. The tools are clunky, and the visualizations difficult to read.Profiling has long been an esoteric skill that only few people ever took the time to learn. The tools are clunky, and the visualizations difficult to read. However, at it's core, performance engineering is very methodical, and follows a lot of rules. AI has also gotten very good at writting code, so in an abundance of luck, they are a match made in heaven!
In this session Frederic will walk through the incredible integration between AI and Profiling and show how performance engineering is now accessible to every engineer regardless of proficiency.read more
Frederic BranczykFounder/CEO, Dash0
☆
Thu 10/1 · 12:30 PM - 12:55 PM
Sovereign AI at Frontier Scale
● Lunch Sessions
Breakout Session
📍 Theater 2
Frontier AI development is increasingly concentrated among a small number of centralized labs, leaving many countries, industries, institutions, and communities dependent on models they cannot fully inspect, adapt, govern, or control.Frontier AI development is increasingly concentrated among a small number of centralized labs, leaving many countries, industries, institutions, and communities dependent on models they cannot fully inspect, adapt, govern, or control. Project Tapestry, led by the AI Alliance, offers a different path: a global, open, consortium-based approach to developing frontier foundation models where partners contribute data, compute, and talent while retaining control of their own data and the ability to build sovereign derivatives aligned to local laws, languages, cultures, industries, and priorities.
In this session, attendees will learn how Project Tapestry re-frames sovereignty as a performance strategy, why open weights alone are not enough, and how collaborative model development can help organizations move from simply consuming AI to actively shaping the foundation models they depend on.read more
Anthony AnnunziataDir of AI Open Innovation, IBM
☆
Thu 10/1 · 12:30 PM - 12:40 PM
The Shared Layer above Your Coding Agents
● Lunch Sessions
Lightning Talk
📍 Theater 3
Developers now have many capable coding agents, but each lives inside its own harness, interfaces, sessions, and control model. Omnigent provides a shared layer across them.Developers now have many capable coding agents, but each lives inside its own harness, interfaces, sessions, and control model. Omnigent provides a shared layer across them. It lets developers compose work across existing agents, move between CLI, web, and mobile while preserving and sharing sessions, and apply contextual policies across agent activity. This talk introduces the meta-harness and explains how Omnigent turns a collection of separate agents into one coherent system.read more
Ankit YadavSolutions Architect, Databricks
☆
Thu 10/1 · 12:30 PM - 12:55 PM
Your Agents Are Already Running. Accounting for What They Did Isn't
● Lunch Sessions
Breakout Session
📍 Theater 4
Every organization I talk to knows their agents are running right now. Almost none can tell you which tickets those agents touched last quarter, or what any of it cost. That gap doesn't stay small.Every organization I talk to knows their agents are running right now. Almost none can tell you which tickets those agents touched last quarter, or what any of it cost.
That gap doesn't stay small. Every day it goes unaddressed, more unattributed spend and unexplained output pile up behind it, and the system gets harder to audit after the fact than it would have been to design correctly from day one.
This talk makes the case for building attribution into agent architecture now, while it's still a design decision and not a forensic project. Drawing on what it actually takes to tie each agent commit and PR back to the specific ticket driving it, Shams Chauthani, CTO of Tempo, walks through where most attribution attempts break down, what a real accounting layer requires architecturally, and why the teams that solve this now will still be able to answer three questions at once: what did the agents cost last quarter, what work items did they close, and which ones touched the code that broke on Friday.
You'll leave with:
Where "AI visibility" tooling quietly becomes guesswork the moment more than one agent touches a task
What it actually takes to tie agent commits and PRs to the ticket driving them, not just report on it after the fact
The specific instrumentation points where agent-to-ticket linkage has to be captured, and what breaks when you skip themread more
Shams ChauthaniCTO, Tempo
☆
Thu 10/1 · 12:45 PM - 12:55 PM
The Real Cost of Scaling AI: An Inference Playbook
● Lunch Sessions
Lightning Talk
📍 Theater 3
Training gets the headlines, but inference pays the bills, and for most AI products, it's where margins are won or lost.Training gets the headlines, but inference pays the bills, and for most AI products, it's where margins are won or lost. Today, the vast majority of inference runs on big cloud providers or abstraction platforms, where convenience is high but cost visibility is low. Teams that carefully optimize their models often ship them on infrastructure chosen by default, then discover at scale that their fastest-growing cost is also their least understood.
This session is a hands-on framework for understanding inference unit economics in practice. We'll start with the problem every team eventually faces: usage grows, the invoice grows faster, and nobody can quite explain the relationship between the two, because layers of managed abstraction hide the variables that drive cost.
The core of the talk is a step-by-step method for making those variables visible. We'll break down how to calculate your true cost per request (not just the sticker price), why utilization - not hardware price - is the number that matters most, and how workload shape (batch vs. real-time, bursty vs. steady, token-heavy vs. compute-light) should drive infrastructure choices. And we'll challenge a common assumption along the way: GPUs aren't the only answer. A growing share of production inference runs efficiently on CPU-only servers, and knowing when that trade-off wins is part of the framework.
With the model in hand, we'll map the decision landscape honestly: serverless APIs, managed GPU endpoints, dedicated GPU compute, and CPU-based inference each win in different scenarios, and attendees will get a simple break-even approach for knowing when to move between them.
We'll close with a brief look at how these principles play out on dedicated infrastructure: one example of how teams that own their inference stack turn cost from a mystery into a lever.
Attendees will walk away with a clear AI infra framework, a checklist for auditing their current inference spend, and a clear-eyed way to answer the question every engineering lead eventually gets asked: "why is our AI bill this high, and what are we doing about it?"read more
Closing the Gap Between AI Ideas and Business Impact
● Lunch Sessions
Breakout Session
📍 Theater 4
The first wave of AI adoption was dominated by pilots, chat bots, and POCs, but the time for experimentation is over. Now, businesses are focused on delivering measurable outcomes quickly, securely, and with predictable costs.The first wave of AI adoption was dominated by pilots, chat bots, and POCs, but the time for experimentation is over. Now, businesses are focused on delivering measurable outcomes quickly, securely, and with predictable costs. Operationalizing AI transformation can be intimidating, but IT leaders have been preparing for this moment for years - they just might not know it. In this breakout session, you will learn about how your business can cross the two chasms separating you from achieving your AI transformation goals - the purpose chasm and the production chasm. Learn how to identify the best opportunities for impact and how your existing investments in cloud transformation provide a strong foundation for your AI transformation journey.read more
Jonathan LaCourDir, AI and Cloud Technology, CDW
☆
Thu 10/1 · 1:00 PM - 1:25 PM
Post-training is the new product moat: why the application layer started training its own models
● Lunch Sessions
Mainstage Talk
📍 Theater 1
Two years ago, post-training was a research term and a checkbox on an inference bill.Two years ago, post-training was a research term and a checkbox on an inference bill. Today Cursor, Cognition, Harvey and others ship their own models, built on open weights and sharpened with reinforcement learning - not because they wanted to become model labs, but because prompting a general model was no longer good enough for their products.
This talk traces how that shift happened and what it means for anyone building AI products. I'll cover the four preconditions that made it possible: open-weight models good enough to specialize, verifiable rewards that made RL practical outside a lab, agents that raised the bar from plausible answers to reliable task completion, and the economics of a small custom model beating a frontier API. I'll then look at how the vendor landscape has organized itself - who sells post-training today, and how much of the work each option leaves in your hands.
Drawing on post-training Llama 3 and 4 at Meta, co-founding Eigen AI, and now leading AI training at Nebius, I'll share how we approach this at Nebius: an AI cloud where training tooling - both our own and our partners' - sits directly on the compute, so teams can post-train their own models without stitching together a stack. I'll ground this in a recent customer case and close with three predictions about where post-training goes next.read more
Di JinVP of AI Training, Nebius
☆
Thu 10/1 · 1:00 PM - 1:10 PM
Runtime is Where AI Lives
● Lunch Sessions
Lightning Talk
📍 Theater 3
Your AI stack didn't add a new application. It added a newparticipant.Your AI stack didn't add a new application. It added a newparticipant.
Agents now make calls, invoke tools, read untrusted content, and act on your systems with credentials of their own - and almost all of it happens at Layer 7, in runtime, where your existing controls were never designed to look. AI opened four new doors at once: autonomous agents, prompt injection, the model supply chain, and identity sprawl. Meanwhile attackers do what they've always done - move across identity, apps, APIs, data, and infrastructure without respecting the tool boundaries we bought.
This session makes the case that runtime is becoming the control plane for modern security, and shows what that actually changes day to day: vulnerability management driven by what's exposed and reachable instead of a CVSS backlog; detection built on behavior instead of static signatures; and the shift frompanic patchtorapid protection - reduce risk now, patch on your timeline.
Brian Joe, CPO and co-founder of Impart Security, will walk through the four AI threat vectors, the unified runtime view, and a practical picture of SecOps in the AI era.read more
Brian JoeCo-Founder and CPO, Impart Security
☆
Thu 10/1 · 1:00 PM - 1:25 PM
Scaling AI: Design to Deployment Continuum Validating AI Infrastructure at Scale
● Lunch Sessions
Breakout Session
📍 Theater 5
As demand for AI-driven solutions grows at an unprecedented scale, the technology industry must overcome complex scalability challenges at every layer, from chiplets to massive AI clusters.As demand for AI-driven solutions grows at an unprecedented scale, the technology industry must overcome complex scalability challenges at every layer, from chiplets to massive AI clusters. This presentation introduces "ScaleX" - a comprehensive framework defining the dimensions of scale: scale-in, scale-up, scale-out, scale-across, and scale-above. We will explore the critical infrastructure hurdles currently facing the industry and examine the diverse architectural approaches being deployed to solve them.
Building high-performing AI applications and agentic systems requires looking beyond the foundational model. It demands optimized inference workflows, network and storage architectures that prevent token-throughput bottlenecks, and rigorous security guardrails that balance efficacy with performance. This session demonstrates how advanced tooling and telemetry can unify these requirements into a single, data-driven feedback loop. Attendees will learn how to refine prompt engineering while serving reliable inference, simulate real GPU memory and storage access patterns to minimize latency, and stress-test security layers against adversarial attacks and data leakage. Ultimately, this presentation provides a path for engineering AI applications that are inherently fast, cost-efficient, and secure.read more
Venkat PullelaCTO, AI Infrastructure, Keysight
☆
Thu 10/1 · 1:00 PM - 1:25 PM
Two Halves of Inference to Build, Eval & Run Agents
● Lunch Sessions
Breakout Session
📍 Theater 2
The AI stack is changing, but whether you’re building an agent or fine-tuning an LLM, the same problem remains: you need the right context to build, evaluate, and deploy reliable AI systems.The AI stack is changing, but whether you’re building an agent or fine-tuning an LLM, the same problem remains: you need the right context to build, evaluate, and deploy reliable AI systems.
Many failures that look like model or prompt problems are actually context problems, with agents making confident decisions that are inaccurate or based on stale data.
Evals have the same context problem. If you evaluate an agent or fine-tuned LLM against non-realistic or incomplete context, you can’t know how it will perform on the tasks and data it will encounter in production.
In this hands-on presentation, Chalk Co-Founder Elliot Marx will show how a unified context layer solves both. A context layer is the missing decision infrastructure to give live, governed data to agents and models at inference time - querying your data where it already lives, so every AI decision gets the real-time context it needs. The same system powers testing and evals, so you can prove that your agent is production-ready.
With a context layer, you turn data that helps AI understand the world it’s acting on into agent context at the moment of inference - and run the full AI decision system. He’ll demonstrate the full workflow across use cases, from fine-tuning an open-weight LLM and serving it in your own cloud to testing an agent in a simulated time-bound environment with knowledge-cutoff data and tool access, evaluating against real-world tasks, and deploying in a sandbox.
Finally, Elliot will share practical lessons from building context and infrastructure for ML and agentic systems used by companies like Whatnot, Turo, and Mission Lane, and a framework for making context and evals a core part of how you build AI.read more
Elliot MarxCo-Founder, Chalk
☆
Thu 10/1 · 1:15 PM - 1:25 PM
The Last Agent Architecture
● Lunch Sessions
Lightning Talk
📍 Theater 3
Models change monthly. Prompts change weekly. The need to re-architect your agent stack feels like a quarterly necessity. Change is inevitable - so build like it.Models change monthly. Prompts change weekly. The need to re-architect your agent stack feels like a quarterly necessity. Change is inevitable - so build like it. This talk explains how to set up a sovereign stack that sticks - flexible primitives, deep observability, and an execution layer that allows for continuous change.read more
Mitchell AldersonSenior Content Engineer, Inngest
Built for you by(617)-PATRICK, Champion AI speaker
Voice AI is the fastest path to ROI in the contact center because it hits your most costly channel on day 1. But the return is architected, not automatic.Voice AI is the fastest path to ROI in the contact center because it hits your most costly channel on day 1. But the return is architected, not automatic. It comes from two early decisions: pointing the agent at the right high-volume, measurable conversations to start, and getting those conversations to actually feel human, so callers stay engaged instead of asking for a person. Get that right and the results follow, the way Amazon Ring's did, ramping from a small slice of calls to their full frontline volume in just 2 weeks. The teams that win don't stumble into ROI. They design for it, starting with the very first call.read more
Dhwani TrivediAgent Engineer, Vapi
☆
Thu 10/1 · 1:30 PM - 1:55 PM
Rebuilding Search for the Agentic Web
● Lunch Sessions
Breakout Session
📍 Theater 5
Abstract: The internet is splitting into two layers: one for humans and one for machine intelligence. This shift creates a new search paradigm, where more compute expands how agents discover, reason, and act.Abstract: The internet is splitting into two layers: one for humans and one for machine intelligence. This shift creates a new search paradigm, where more compute expands how agents discover, reason, and act. Learn what the next layer of the agentic web requires.
Outline:
What Tavily is: the web access layer for AI agents
History of the web: humans -> rise of agents
The internet is splitting into two layers: one optimized for human attention, another for machine intelligence
Search becomes an iterative process: agents search, read, identify gaps, verify claims, and search again
More compute means better search
Importance of agent harness. Retrieval, tools, memory, permissions, and evaluation determine what an agent can actually do
Search is part of the learning loop: agents can learn which sources to trust, when to corroborate, and when to abstain
The next layer of the agentic web: reliable, structured, source-backed web access that lets agents observe, reason, act, and improveread more
Rotem WeissFounder/CEO, Tavily
☆
Thu 10/1 · 1:30 PM - 1:55 PM
The Agentic OS: a first-principles approach to building agents that get things done
● Lunch Sessions
Breakout Session
📍 Theater 2
AI agents are remarkable. They've given us the ability to reason over complex, multi-dimensional knowledge and capabilities across an organization, and take actions at a scale that previously demanded far more time and resources.AI agents are remarkable. They've given us the ability to reason over complex, multi-dimensional knowledge and capabilities across an organization, and take actions at a scale that previously demanded far more time and resources. There is a meaningful gap, though, between agents that demo well (AI always demos well) and agents that actually move the needle.
As the agentic discourse moves from models and frameworks to scaled deployment and ROI, the harder question is whether agents can survive contact with a real organization: real data, fragmented systems, ambiguous outcomes, and decisions that carry actual risk.
This talk draws from our experience at Postman building and operating a family of 30+ AI agents in production across our enterprise organization. Integrations multiplied, context fragmented, and the emerging mesh of agents calling one another began to resemble the distributed-system architectures many of us spent the last decade learning to avoid. AI-native is not just about deploying more agents, and much more about making the business legible to them: domains, judgment, strategy, permissions, and outcomes.read more
Ankit SobtiCo-Founder, Postman
☆
Thu 10/1 · 1:30 PM - 1:55 PM
What Trusting an Agent With 65,000 Files Taught Us
● Lunch Sessions
Mainstage Talk
📍 Theater 1
For most people, AI assistants start with finding and summarizing, and in the usage we see across AI tools connected to a shared file store, that is where most of it stays: the majority of requests are retrieval, and about...For most people, AI assistants start with finding and summarizing, and in the usage we see across AI tools connected to a shared file store, that is where most of it stays: the majority of requests are retrieval, and about half of users ever take an action on a file. Acting on real, shared work is a different bar: reorganizing a folder a colleague depends on, renaming a thousand files to a convention, assembling a package for a client. In our experience, the gap between an assistant that suggests and one that finishes is rarely model quality. It is trust, and trust can be designed in.
This session reports what we learned building an agent that operates on shared, permissioned content at scale: nine organizations running it on their own files for months, including a 65,000-file archive, and side-by-side tests against general-purpose chat agents. It lays out what changed the behavior of the people who moved from asking to acting.
What the audience will take away:
- The trust ladder. Adoption starts with low-risk, reversible actions (organize, rename, convert) and climbs from there. Skipping rungs loses people.
- Five controls before an agent touches shared files: a plan the user can read and edit; write access granted per scope and enforced server-side, never by the model; rollback for every change, verified against the actual file state; a record of what ran, when, and on whose behalf; and results that land where the team already works, in sync, rather than in a copy.
- Context lives with the content. Why a file system with history, permissions and collaborators is a better substrate for an agent than a chat window, and how to keep that context open to any model and any tool through standard interfaces rather than locking it to one.
We will share what the research changed in the design: why an approval step has to behave the same way every time before people can learn to trust it; why a plan that is complete but unread is a failed plan, and what made people read it; why the language the agent uses while it works matters as much as the result; and why undo moved adoption further than any gain in accuracy. Attendees leave with a checklist for evaluating or building agents that people will let near their work.read more
Vijay RangarajanVP of Product Management, Dropbox
☆
Thu 10/1 · 1:30 PM - 1:55 PM
Your agent pays twice for every search: how we cut token spend by 90+% with better accuracy
● Lunch Sessions
Breakout Session
📍 Theater 4
Every agent that searches the web pays twice: once for the search call, and again for the tokens that search puts into the model's context. The answer is usually a few sentences.Every agent that searches the web pays twice: once for the search call, and again for the tokens that search puts into the model's context. The answer is usually a few sentences. The agent pays for the whole page, and in a multi-step loop it pays for it again on every turn.
We'll show real production traces of where the tokens go and why retrieval, not generation, is the biggest line on most agent bills. Then we'll show what we changed: return the passages a query needs instead of whole pages, tune extraction to the question, and cap how much context the agent can take. We tested this across 12 frontier models and 11,000 blind-graded answers. Token spend fell by more than 90% while accuracy went up.
You get a metric (cost per correct answer, not cost per request), an open calculator to run the numbers on your own traffic, and a five-point checklist for finding token waste in your pipeline.read more
Louis WalshHead of Forward Deployed Engineering, Exa
☆
Thu 10/1 · 1:45 PM - 1:55 PM
The Rise of Open Models: Why Pay for Frontier?
● Lunch Sessions
Lightning Talk
📍 Theater 3
Most engineering orgs default to frontier models for everything.Most engineering orgs default to frontier models for everything. In this session, I'll walk through the routing policy Faros built and validated on hundreds of real engineering tasks, using an execution-free eval framework that works across any codebase without test infrastructure. Open models matched frontier quality on most task types at roughly half the cost. You'll leave knowing which tasks still need frontier models and why. Plus how to run the same analysis on your own work.read more
Every hour spent making infrastructure work for agents is an hour developers could spend making agents more useful.Every hour spent making infrastructure work for agents is an hour developers could spend making agents more useful. A general-purpose VM gives an agent somewhere to run, but leaves the developer responsible for preserving context, securing generated code, coordinating subagents, governing tool access, and tracking costs across services. As agents take on longer and more ambitious tasks, supporting infrastructure becomes a substantial engineering project of its own.
This talk explores computing primitives designed around how agents actually work: bursts of execution, waits for models and tools, and sessions that pause, branch, and continue. We will go beneath the APIs to examine hardware virtualization with Firecracker microVMs, the roles of Linux namespaces and cgroups, and how CPU accounting supports billing for actual consumption. We will then connect those foundations to reusable environment templates, swarms of agents, governed actions, and observability that turns execution traces into signals for improvement. Using DigitalOcean Managed Agents as the concrete example, we will show how bringing execution, inference, storage, and tool consumption onto a single bill helps teams understand the cost of successful work and spend more engineering time improving it.read more
Salman ParachaSVP, AI, DigitalOcean
☆
Thu 10/1 · 2:00 PM - 2:25 PM
From Recommendations to Agents: Building Goal-Driven Consumer AI at DoorDash
● AI Builders
📍 Theater 2
Modern consumer AI is moving beyond one-shot predictions toward systems that actively
reason, plan, and adapt.Modern consumer AI is moving beyond one-shot predictions toward systems that actively
reason, plan, and adapt. In this talk, Sudeep explores how personalization is evolving from
static recommender systems into agentic, goal-driven AI experiences that operate across
search, discovery, decision-making, and execution. Drawing from real-world production
systems, the session covers how large language models, deep learning, and
reinforcement-style feedback loops can be combined to build consumer-facing agents that
understand intent, maintain context over time, and take coordinated actions on behalf of
users. Topics include agentic product discovery, multi-step decision flows, trust and
controllability in consumer agents, and practical evaluation strategies for these systems at
scale. Attendees will leave with concrete frameworks for designing AI systems that feel
helpful, adaptive, and human-centered - while remaining reliable, efficient, and grounded in
real business constraints.read more
Sudeep DasHead of AI, New Verticals, DoorDash
☆
Thu 10/1 · 2:00 PM - 2:25 PM
Scaling Enterprise RAG Systems: Distributed Indexing and Dynamic Retrieval for Terabyte-Scale Knowledge Systems
● The AI Stack
📍 Theater 3
Standard vector search architectures collapse under the weight of indexing latency, ballooning storage costs, and degraded retrieval precision when enterprise RAG systems scale to terabyte-scale production knowledge bases, hitting a Scaling Wall.Standard vector search architectures collapse under the weight of indexing latency, ballooning storage costs, and degraded retrieval precision when enterprise RAG systems scale to terabyte-scale production knowledge bases, hitting a Scaling Wall.
This session explores the engineering patterns required to build a high-throughput enterprise search infrastructure. Using our production implementation as a case study, we graduate from basic vector stores to a distributed indexing strategy. We will analyze the critical trade-offs between Approximate Nearest Neighbor (ANN) accuracy and memory overhead, the impact of hybrid metadata-vector filtering on query latency, and the necessity of partitioned retrieval architectures to maintain sub-second response times.
Attendees will gain a blueprint for a "Knowledge Lake" architecture, learning how to handle incremental indexing at scale, manage vector infrastructure costs, and implement dynamic query routing to ensure system consistency across distributed environments.
Key Takeaways:
* Distributed Indexing & Sharding: Partition billion-vector datasets across distributed nodes using HNSW/IVF algorithms to ensure sub-second retrieval without exhausting cluster memory.
* Trade-off - ANN accuracy vs Memory: Tune ANN indexes for billions of vectors, balancing retrieval speed against infrastructure spend with tiered storage.
* Partitioned Retrieval & IVF Strategies: Engineering strategies for segmenting vector space using Inverted File Indexing (IVF), centroid-based sharding, and cluster-level pruning to achieve high search completeness with minimal compute overhead.
* Hybrid Filtering & Routing: Implementation patterns for combining structured metadata filters with vector similarity to optimize retrieval performance.
* Operational Consistency: Monitoring ingestion lag, managing incremental indexing, and ensuring data integrity across distributed vectors.
Engineering Trade-offs & Operational Insights:
* Ingestion Scaling vs. Index Freshness: Dynamic task mapping in Airflow and Spark to parallelize the embedding bottleneck while managing vector database write-limits.
* Vector Storage Economics: Reducing high-dimensionality embeddings cost through product quantization (PQ) and binary hashing.
* System Resiliency: Handling retrieval hot spots and implementing circuit breakers for bulk data backfills that threaten to saturate vector endpoints.read more
Yuri ShelkeDistinguished Engineer, Capital One
☆
Thu 10/1 · 2:00 PM - 2:25 PM
Securing Enterprise and Sovereign AI at Runtime
● AI Strategy
📍 Theater 5
Enterprise and sovereign AI are already moving into production. Agents are retrieving sensitive data, calling models and tools, taking actions, and generating detailed traces.Enterprise and sovereign AI are already moving into production. Agents are retrieving sensitive data, calling models and tools, taking actions, and generating detailed traces. Traditional security protects data at rest and in transit, but it leaves a gap as sensitive information moves through prompts, agents, models, tools, and regions.
This session looks at how Runtime AI Data Control addresses that gap. You’ll learn where sensitive data shows up across agent workflows, why agent traces require a different approach than traditional logs, and how to protect sensitive data without limiting what AI systems can do or where they can operate.
We’ll also share lessons from real deployments, common failure points, and a practical way to assess whether enterprise and sovereign AI systems are ready to scale securely.
Key talking points:
* Where sensitive data shows up across an agent workflow, from retrieval and prompts to tool calls, memory, outputs, and traces
* Why an agent trace is not the same as a traditional log: a log describes the system, while a trace can describe a person
* How Runtime AI Data Control protects sensitive data across enterprise and sovereign AI deployments without limiting what AI systems can do
* What it takes to scale enterprise and sovereign AI, including common failure points and practical lessons from real deploymentsread more
Amruta MoktaliChief Product, Skyflow
☆
Thu 10/1 · 2:00 PM - 2:25 PM
The New Economics of AI.
● AI Frontiers
📍 Theater 1
This talk builds on our State of AI report released in December 2025, one of the largest empirical studies of real-world LLM usage to date, analyzing 100 trillion tokens of inference data from the OpenRouter platform.This talk builds on our State of AI report released in December 2025, one of the largest empirical studies of real-world LLM usage to date, analyzing 100 trillion tokens of inference data from the OpenRouter platform. Rather than relying on benchmarks, we examine how models are actually used in practice across tasks, geographies, and time. Among the original findings: open-weight models now serve roughly a third of all tokens in an increasingly fragmented ecosystem, creative roleplay is the dominant open-source use case, and programming is the fastest-growing category overall. We'll revisit these findings with fresh data and updated trends, including the continued rise of agentic inference (multi-step, tool-integrated workflows) alongside cost-usage dynamics, geographic shifts, model retention patterns, and what these evolving patterns mean for model builders, infrastructure operators, and investors.read more
Beyond the Benchmark: Building for the Complexity of Real Marketplaces
● Applied AI
📍 Theater 4
Most AI research is evaluated against static benchmarks with clean labels, single-language inputs, and well-defined reward functions.Most AI research is evaluated against static benchmarks with clean labels, single-language inputs, and well-defined reward functions. But what happens when those models meet a live, global, two-sided marketplace, where hiring signals are sparse, feedback is delayed by weeks and queries span 30+ languages? The difference between a good recommendation and a bad one has a real impact on outcomes for all players in a marketplace.
This talk walks through what it actually takes to build a robust, end-to-end AI experience for a global talent marketplace. It presents five research contributions from Upwork: four that each solve a critical layer of that stack, and a fifth that tackles how to evaluate the AI agents increasingly operating within it.
It starts with matching. GraphMatch fuses language models with graph neural networks to capture the relational structure (who's worked with whom, how skills connect, how the marketplace evolves) that text embeddings alone miss. But finding the right candidates is only the beginning. They need to be ranked for hiring intent, not clicks, which is what Rank-Aware Indigo-DPO delivers, using preference optimization across 30+ languages to align search results with actual hiring outcomes. Even with strong matching and ranking, marketplace feedback arrives late: a hire confirmed weeks after a search session is still a valuable signal. IF-DFM uses influence functions to keep conversion models fresh without retraining, closing up to 85% of the gap to an oracle with perfect labels. And all of this has to run at production speed. DispatchQA demonstrates that a fine-tuned 8B parameter model matches GPT-4o on search function-calling at half the latency.
Separately, as AI agents increasingly participate in marketplaces like Upwork, not just powering the platform but doing the work, we need new ways to evaluate them. UpBench, our fifth contribution, is a dynamically evolving benchmark grounded in hundreds of real Upwork jobs, evaluated by expert freelancers using hand-crafted rubrics. It reveals that today's best agents complete fewer than half of real professional tasks on their first attempt, and that human-in-the-loop feedback rescues roughly one in five failures.
Together, these five papers offer a view from inside a team building AI where the stakes are real, the signals are noisy, and the users are on both sides of every transaction.read more
Jon PappasSr. Dir. Technical Product Management, Upwork
☆
Thu 10/1 · 2:30 PM - 2:55 PM
Harness Engineering: What Changes When Agents Write Most of the Code?
● AI Builders
📍 Theater 2
As coding agents get better, the limiting factor is no longer just model capability.As coding agents get better, the limiting factor is no longer just model capability. It is whether your team has built the right harness around the model: the codebase structure, task boundaries, feedback loops, evals, tests, review processes, and operational guardrails that let agents work effectively over longer time horizons.
This talk will cover the emerging shift from pair programming with AI to agentic delegation. It will examine what kinds of engineering work agents handle well today, where they still break down, and what changes teams are making to get useful results in practice. Topics will include how to make a codebase more agent-friendly, how to scope tasks so agents do not thrash, how to use tests and evals as control systems, and how human review changes when the AI is doing more of the implementation work.
Attendees will leave with a practical framework for thinking about “harness engineering,” along with concrete patterns they can apply to code review, task design, evals, and team workflows as AI systems take on a larger share of software development.read more
Charlie GuoDeveloper Experience Engineer, OpenAI
☆
Thu 10/1 · 2:30 PM - 2:55 PM
Human Safety Was Hard Enough. Now We Have Agents.
● AI Strategy
📍 Theater 5
For the last decade, platform safety meant one thing: keeping humans safe from adversarial humans online. We built systems, policies, and entire product organizations around that assumption. That assumption is now obsolete.For the last decade, platform safety meant one thing: keeping humans safe from adversarial humans online. We built systems, policies, and entire product organizations around that assumption. That assumption is now obsolete.
We are entering a world with three new safety vectors that no platform has fully mapped yet: adversarial humans manipulating agents to cause harm, agents causing harm to humans, and agents in conflict with each other. The platforms sitting in the middle have no rulebook, no established liability framework, and in most cases, no product roadmap that accounts for any of this.
I’ve spent the last five years building Trust & Safety platforms at scale, across content moderation, fraud detection, youth safety, and messaging integrity. The hardest problems I’ve worked on involved adversarial human actors behaving in ways we could anticipate, model, and eventually defend against. Agents are a different category entirely. They can learn, adapt, and cause harm in ways that no policy written today will cleanly cover. And when harm happens, the accountability question becomes genuinely unsolved: is it the platform? The agent? The company that built the model? The user who trained it? Or all of the above?
This talk maps the new safety surface that product and platform leaders need to start designing for now, while the field is still early enough to shape it thoughtfully. We will walk through each harm vector, the accountability gaps each one creates, and the product and governance questions leaders need to be asking today.
Attendees will leave with an understanding of the emerging agent safety surface, a framework for thinking through accountability across the four-way responsibility stack, and the specific product questions they should be bringing into roadmap conversations right now.read more
Vidya SrinivasanProduct Staff, Meta
☆
Thu 10/1 · 2:30 PM - 2:55 PM
Rethinking Inference for the Agent Era
● The AI Stack
📍 Theater 3
Agents have changed the economics of AI inference. A chatbot’s cost scales roughly linearly with the number of requests; an agent’s scales multiplicatively.Agents have changed the economics of AI inference. A chatbot’s cost scales roughly linearly with the number of requests; an agent’s scales multiplicatively. A single task can fan out into hundreds of model calls, each carrying a repeated context prefix and adding latency that compounds across tool calls and reasoning steps. As open-weight models keep improving and agentic workloads grow, this shift exposes the limits of traditional request-level optimization. Inference infrastructure becomes a first-class concern, one that often shapes performance and cost as much as the model itself. In this talk, we explore what changes when you optimize for the whole task rather than the individual request, and how FriendliAI is rethinking the inference cloud for the era of agentic AI.read more
Gon ChunCEO, FriendliAI
☆
Thu 10/1 · 2:30 PM - 2:55 PM
The Million-Dollar Pilot Trap: Why Most AI Agents Never Scale
● AI Frontiers
📍 Theater 1
AI experimentation is everywhere. Enterprise-scale AI deployment is not. While organizations race to adopt AI agents, many remain trapped in pilots that never deliver meaningful business value. The challenge isn't access to AI.AI experimentation is everywhere. Enterprise-scale AI deployment is not. While organizations race to adopt AI agents, many remain trapped in pilots that never deliver meaningful business value. The challenge isn't access to AI. It's building the operational foundation required to deploy AI securely, govern it effectively, and scale it across the enterprise. What can we learn from enterprises already running AI agents in production across millions of customer interactions? Dive into the realities of operationalizing AI, the governance models that build trust, and the measurable outcomes organizations are achieving today.read more
5 Teams, 5 Different AI Strategies: Here's What Worked
● AI Strategy
📍 Theater 5
Mollie BreenCo-founder & CEO, OverAI
Ernesto MoralesVP, Information Technology, International Materials
☆
Thu 10/1 · 3:00 PM - 3:25 PM
From Policy to Practice: Scaling Responsible AI Inside a Global Bank
● AI Frontiers
📍 Theater 1
Terah Lyons has seen AI governance from both sides - as the architect of the first U.S.Terah Lyons has seen AI governance from both sides - as the architect of the first U.S. national AI policy strategy and founding CEO of the Partnership on AI, and now as the person responsible for turning policy into practice inside one of the world's largest financial institutions. In this fireside chat with Washington Post tech reporter Gerrit De Vynck, Lyons shares concrete lessons from deploying AI at scale inside JPMorganChase - what responsible adoption actually looks like under real regulatory and operational constraints, where global governance frameworks help versus where they slow real execution, and what it takes to balance innovation with safety when the stakes, and the scrutiny, are this high.read more
Terah LyonsGlobal Head of AI & Data Policy, JPMorganChase
Gerrit De VynckTech Reporter, Washington Post
☆
Thu 10/1 · 3:00 PM - 3:25 PM
Our Agent Resolves Production Bugs in Minutes. Here's What We Learned
● Applied AI
📍 Theater 4
At Navan, production issues accumulate in Jira while engineers manually trace errors across three disconnected systems - the ticket description, New Relic observability data, and the responsible codebase - before they can even begin fixing anything.At Navan, production issues accumulate in Jira while engineers manually trace errors across three disconnected systems - the ticket description, New Relic observability data, and the responsible codebase - before they can even begin fixing anything. With 120+ microservices, this triage-to-resolution cycle averaged 4.2 hours per ticket. Roberto Milev and Uday Kanagala (Navan) built an agent that does it in 9 minutes.
Sherlock is an end-to-end autonomous agent that resolves production issues by orchestrating across Jira, New Relic, and GitHub. Given a Jira ticket - whether it contains a precise exception message or just a screenshot of a generic error page - the agent autonomously traces the error through production stack traces, identifies the responsible service and code, generates a fix, and opens a pull request. It runs on Claude SDK deployed as Docker containers on AWS AgentCore runtime, with MCP tool servers providing authenticated access to each system.
The hardest engineering problem was observability-driven error tracing. Tickets vary wildly in specificity. The agent extracts candidate error signatures from ticket content (including screenshots), formulates NRQL queries against New Relic, and uses time-window scoping - the ticket's creation timestamp and temporal references in the description - to constrain the search across 120+ services. It iteratively refines queries until it converges on matching error occurrences, retrieves full stack traces, and maps them to specific files and line numbers. This transforms a vague report like "recommendations aren't loading" into a precise pointer to the failing method.
A critical design decision: the agent operates on a full local clone of the repository, not API-based file access. It reads the failing class, its imports, dependent services, configuration files, and existing test suites. It also ingests agentic context files (CLAUDE.md) that describe business logic and conventions - giving it the same onboarding context a new engineer would receive. This produces convention-aware fixes that follow the repository's existing patterns.
Not every ticket can be fully resolved autonomously. Sherlock is designed for graceful degradation: it goes as far as available information allows, documents its progress, and pauses when it needs human input. The workspace - cloned repo, analysis state, session context - is persisted across AgentCore Memory, S3, and ECS. When a human replies with direction days later, the agent resumes exactly where it left off.
Evaluated on 100 production tickets: 41% fully resolved (PR created), 38% partially resolved (diagnosis posted to ticket), 9-minute average time-to-PR, 68% PR acceptance rate after human review, and 72% error recovery rate when initial tool calls failed.
Key takeaways:
- Architecture for safe autonomous debugging across disconnected enterprise systems
- When agents should stop and ask for help - designing graceful degradation into production agent systems
- Practical patterns for deploying agents that interact with real production infrastructureread more
Roberto MilevChief Architect, Navan
☆
Thu 10/1 · 3:00 PM - 3:25 PM
Post-Training LLMs for Entertainment: Improving Engagement and Safety with Non-Verifiable Rewards
● The AI Stack
📍 Theater 3
Training LLMs for entertainment applications, such as AI characters and interactive storytelling, presents a unique opportunity: how to develop models that are engaging and safe.Training LLMs for entertainment applications, such as AI characters and interactive storytelling, presents a unique opportunity: how to develop models that are engaging and safe. Measuring model responses is a specialization where there is not merely a set universe of objective criteria, as in other fields. As a result, we often use signals from user behavior, such as swipe interactions and whether a user continues the conversation or disengages, to infer preferences and guide model improvements.
This talk will explore how we approach post-training in this setting using reinforcement learning with real-world user interaction data. We will discuss how to identify which user behaviors are meaningful and how to use them to improve models despite the lack of a clear ground truth.
We will share practical techniques for balancing multiple goals. This includes methods for structuring training so the model learns from both positive signals (e.g., continued engagement) and negative signals (e.g., undesirable outputs), as well as strategies for prioritizing different objectives at different stages of training.
We will also cover how testing and training models on difficult prompts helps uncover weaknesses and improve reliability in real-world use, and how improvements discovered in larger models can be transferred to smaller, more efficient models without sacrificing performance.
Grounded in real-world implementation, this session will provide a clear framework for training and improving LLMs in domains where success is difficult to measure directly. Attendees will leave with practical approaches to working with imperfect user feedback, balancing competing objectives, and applying these methods to production AI systems.read more
Manqing LiuResearch Engineer, Character.ai
☆
Thu 10/1 · 3:00 PM - 3:25 PM
True Zero Data Retention Is Structural, Not a Clause in a Contract
● AI Builders
📍 Theater 2
Every model that touches the web inherits the privacy policy of its search provider. This means Zero Data Retention is only enforceable up to your search vendor's boundary, unless it's a property of a vertically integrated stack.Every model that touches the web inherits the privacy policy of its search provider. This means Zero Data Retention is only enforceable up to your search vendor's boundary, unless it's a property of a vertically integrated stack. If they're calling an external index for some queries, then you can't guarantee true ZDR for your users. In this talk, you'll get practical advice for how to trace the retrieval path your queries actually take; how to verify end-to-end privacy; why agents are compounding this privacy gap; and three questions to take back to your team to ensure reliably enforceable ZDR across all endpoints without degradation in result quality.read more
Carlos AcevedoSr Dir, Business Development, Brave
Built for you by(617)-PATRICK, Champion AI speaker
Interactive agents aren’t scaling. Every company is coping with runaway costs, lack of governance, and difficulty automating.Interactive agents aren’t scaling. Every company is coping with runaway costs, lack of governance, and difficulty automating.
This has led to the need for software factories: a closed-loop system for automating development, with primitives that provably improve throughput and efficiency over time.
Right now, folks are trying to roll their own internal software factories from bare metal, solving the same problems 100 different ways. We need shared, flexible infrastructure that everyone can build upon.
In this talk, we’ll share the core building blocks every company needs to build their own self-improving software factories based on our own experience building Warp Factories:
- Cloud infrastructure to centralize agent conversations, skills, and security permissions
- Configuration as code for humans and agents to manage that cloud infrastructure
- Metrics for humans and agents to track efficiency and code quality, beyond “number of PRs shipped”
- Self-improvement loops for agents to review past conversations, score the quality, and propose improvements to the skills driving the factory
You’ll walk away with a clear understanding of how to improve coding agent ROI, and how to deploy a system of your own using Warp Factories.read more
Zach LloydFounder/CEO, Warp
☆
Thu 10/1 · 4:00 PM - 4:25 PM
Premium Inference for the Agentic AI Era
● AI Builders
📍 Theater 2
As AI shifts from single-turn prompts to autonomous, multi-step agents, inference speed is becoming foundational to real-world adoption. Agents must reason, retrieve, plan, call tools, and respond across complex workflows - often requiring many model interactions in sequence.As AI shifts from single-turn prompts to autonomous, multi-step agents, inference speed is becoming foundational to real-world adoption. Agents must reason, retrieve, plan, call tools, and respond across complex workflows - often requiring many model interactions in sequence. This session explores why high-performance inference is critical for scalable agentic AI, and how purpose-built infrastructure can reduce latency, improve responsiveness, and lower operational cost. Featuring perspectives from General Compute, the discussion will highlight how next-generation inference clouds are helping bring faster, more practical AI systems into production.read more
Vasanth MohanSr Dir of Product Marketing and Developer Relations, SambaNova
Jason GoodisonCTO, General Compute
☆
Thu 10/1 · 4:00 PM - 4:25 PM
Right Query, Wrong Answer: Why AI Analytics Is So Hard, and How to Fix It
● The AI Stack
📍 Theater 3
Every product team has an analysis backlog. Questions pile up, same metrics get re-sliced weekly, and when answers finally arrive, validation is manual and inconsistent. At many companies, the ratio is 80% execution mechanics to 20% actual thinking.Every product team has an analysis backlog. Questions pile up, same metrics get re-sliced weekly, and when answers finally arrive, validation is manual and inconsistent. At many companies, the ratio is 80% execution mechanics to 20% actual thinking. We wanted the inverse.
So we built AI Analyst - an open-source, multi-agent system powered by Claude Code that automates the full analytical workflow: from a business question to a validated, stakeholder-ready slide deck in roughly 30 minutes instead of days and weeks.
What We built:
The system orchestrates 18 specialized agents across a 4-phase pipeline - Frame, Analyze, Story, Deck - using DAG-based parallel execution. A Question Framing agent turns vague asks into testable hypotheses. Eight analysis agents run concurrently (descriptive, cohort, root cause, trend, validation, opportunity sizing). A Story Architect designs narrative beats before any chart is created - solving the "chart looking for a story" problem. A Chart Maker generates Storytelling-with-Data-styled visualizations with collision detection. Every output passes through 4-layer validation: structural, logical, business rules, and Simpson's Paradox checks, graded A through F.
The system includes a knowledge layer that compounds over time: query archaeology captures proven SQL patterns, a corrections log ensures mistakes are never repeated, and organization-specific context (metrics definitions, business rules, team glossary) accumulates with use. It connects to any data source - CSV, DuckDB, Postgres, BigQuery, Snowflake.
What I learned:
The hardest part wasn't building the agents - it was getting an organization to trust and adopt AI-assisted analytics. Five real blockers emerged: security and legal approvals, data classification concerns, warehouse access, API cost ownership, and analyst skepticism. I'll share specific tactics for each, including the one that worked best: "show the output, not the pipeline."
The critical design principle: AI can execute analysis fast, but it cannot tell you what question to ask or whether the answer makes business sense. We designed the system around human validation, not human replacement. The learning flywheel - run, review, spot errors, correct, store, improve - is what makes the system get meaningfully better over time.
This talk is for data leaders, product managers, and engineers who want to democratize analytics without sacrificing rigor.read more
Hai GuanHead of Data, Ontra
☆
Thu 10/1 · 4:00 PM - 4:50 PM
The State of AI Regulation: Policy, Law, and Industry
● AI Frontiers
📍 Theater 1
AI regulation in the U.S. is taking shape in real time, with states moving quickly as federal agencies, courts, and industry work to keep pace.AI regulation in the U.S. is taking shape in real time, with states moving quickly as federal agencies, courts, and industry work to keep pace. This panel brings together leaders shaping AI law, policy, and governance to examine where regulation is working, where gaps remain, and what comes next. Moderated by Washington Post tech reporter Gerrit De Vynck, the conversation will explore the tension between state and federal approaches, innovation and oversight, and what a workable regulatory framework could look like as AI capabilities continue to advance.read more
Andrew FreedmanCo-Founder, CEO, Fathom
Gerrit De VynckTech Reporter, Washington Post
Kevin FrazierProfessor, AI Innovation and Law Program, University of Texas School of Law
Rebecca Bauer-KahanCalifornia State Assemblymember, District 16, California State Assembly
☆
Thu 10/1 · 4:00 PM - 4:25 PM
Transform Twice or Get Left Behind
● AI Strategy
📍 Theater 5
There’s a growing trend of companies that implement AI…. and fail to see any real impact for their considerable investment.There’s a growing trend of companies that implement AI…. and fail to see any real impact for their considerable investment. It’s relatively easy to provide access to technology - but creating durable organizational change and accelerating work is a larger (and harder) endeavor.
In an AI-first market, if you’re only working on one transformation, you’re already behind. This talk provides an overview into the two types of fundamental transformation and reinvention modern technology companies must undertake to compete and remain relevant.
The first transformation is inward - changing how the organization actually operates. Reinventing processes, systems, and the assumptions underneath them, rather than layering AI on top of workflows designed for a pre-AI world. AI is an amplifier: point it at a team with a shared standard and apply creativity and you get reinvention; point it at one without it and you just multiply chaos.
The second transformation is outward - what you bring to market. Agentic AI has reset the bar for every software product. Winners now need three ingredients to get things right: the context for how work actually connects, the intent behind what a user is asking for, and the judgment to act on both. Provide one without the others and the magic never shows up
Most organizations are working through one of these transformations, or neither. The ones pulling ahead are doing both at once - and treating them as a single, compounding reinvention.
This session offers a diagnostic framework leaders can apply to their own organization next week: which transformation are you neglecting, what happens when you skip either one, and how the two reinforce each other. We'll ground it in the honest, still-unfinished story of Smartsheet - a platform company working through both transformations in parallel - what’s working, what comes next, and why the journey is far from over.
You'll leave being able to name where your own transformation is stuck - and what it may cost to stay thereread more
Janelle MilodragovichVP, AI Market Strategy, Smartsheet
This talk introduces a specific, implementable architecture for context-aware agent authorization - using structured memory as a first-class input to the policy evaluation hot path.This talk introduces a specific, implementable architecture for context-aware agent authorization - using structured memory as a first-class input to the policy evaluation hot path.
The core argument is that agent tool calls occupy a gap in the authorization model taxonomy. No single existing model handles them well:
RBAC can confirm a user has a role that includes Jira access, but cannot scope a tool call to a specific project without creating a role per project.
ABAC can evaluate request parameters against static constraints, but requires someone to have pre-configured those constraints - it doesn't know why project ALPHA is the right scope.
ReBAC can traverse a relationship graph to determine that a user is related to project ALPHA through a team, but doesn't account for temporal context (the user was on ALPHA last quarter, now on BETA).
The missing input is situational context: what is this user currently working on? What has their agent been doing in this session? What tools do people in this user's role typically use?
The talk will present:
A memory architecture designed for authorization - not general-purpose agent memory. Three typed stores (worker memory from identity/HR/project systems, agent session memory from conversation context, organizational baseline memory from usage analytics), each with different trust levels, update semantics, and confidence scoring. Temporal validity ranges on every fact for point-in-time audit queries. This draws on and synthesizes ideas from the academic agent memory literature (Zep/Graphiti for temporal knowledge graphs, Hindsight for belief evolution with confidence, MIRIX for typed multi-agent stores) but is purpose-built for the authorization use case.
A write/read split architecture for the hot path. Memory facts are extracted asynchronously (tolerating LLM latency for fact extraction and confidence scoring). Authorization decisions are evaluated synchronously against pre-compiled facts using Datalog - deterministic, terminating, auditable, sub-millisecond. The evaluation does not call an LLM. It does not make network round-trips to external stores. The facts are local and pre-compiled.
Memory-informed delegation policies. Concrete Datalog examples showing how a delegation can reference memory predicates instead of static parameter constraints. A rule that says "allow this Jira create_issue call if the target project is one the user is currently working on, per memory, with confidence above 0.7" - and if confidence is below the threshold, escalate to a human approval flow. Plus anomaly detection via organizational memory baselines: "this tool is not typical for this user's role" as a governance signal, not a hard deny.
An honest treatment of the unsolved problems - particularly the local authentication and configuration barrier for non-technical users, and the challenge of governing memory itself (provenance, retention, review cycles) when memory facts influence authorization decisions.read more
Robert ChiniquySoftware Engineer, C1
☆
Thu 10/1 · 4:30 PM - 4:55 PM
Beyond the Demo: Shipping Computer-Use Agents in Production (What Breaks, What Scales, and Why)
● The AI Stack
📍 Theater 3
Computer-use agents, AI systems that interact directly with browsers and graphical user interfaces, have rapidly moved from impressive demos to early production deployments.Computer-use agents, AI systems that interact directly with browsers and graphical user interfaces, have rapidly moved from impressive demos to early production deployments. However, operating these agents reliably in real environments introduces a class of challenges that are often underestimated or ignored in early implementations.
This talk presents a practical, experience-driven examination of computer-use agents deployed in live business workflows. It traces the evolution from low-level browser control approaches (e.g., DOM- and vision-based interaction) to higher-level abstractions and orchestration layers, highlighting the architectural tradeoffs that emerged only under real-world load, UI drift, and partial system failures.
Key topics include:
Common failure modes in computer-use agents, including UI volatility, non-deterministic model behavior, timing and synchronization issues, and brittle task decomposition
Reliability strategies such as constrained action spaces, verification loops, retries and recovery, fallback paths, and human-in-the-loop controls
Observability gaps unique to UI-driven agents and practical techniques for debugging, monitoring, and post-incident analysis
Latency, cost, and throughput tradeoffs when scaling computer-use agents beyond isolated tasks
Security, compliance, and blast-radius considerations when agents operate on live systems
The talk also challenges the assumption that increased autonomy always leads to better outcomes, and proposes a more pragmatic model for deploying computer-use agents as constrained, inspectable systems rather than fully autonomous actors.
Attendees will leave with a concrete mental model for evaluating when computer-use agents are appropriate, architectural patterns that improve reliability, and a set of anti-patterns to avoid when moving from demo to production.read more
Abbas RazaCTO, ReFocus AI
☆
Thu 10/1 · 4:30 PM - 4:55 PM
Beyond the Demo: Validating Demand for AI Products
● Applied AI
📍 Theater 4
An impressive AI demo can attract attention, but what signals show that a product is worth building and scaling?An impressive AI demo can attract attention, but what signals show that a product is worth building and scaling? This talk explores how to distinguish initial enthusiasm from meaningful demand, test whether AI fits a real user workflow, and assess repeat usage, willingness to pay, and delivery costs. Drawing on experience in search, advertising, and fintech product management, Victoria will share a practical approach to deciding what to validate before expanding investment. Attendees will leave with key questions to guide early product decisions.read more
The AI-Pilled Organization: Why Individuals Are Accelerating and Institutions Aren't
● AI Strategy
📍 Theater 5
Over the past year I've spent time inside more than 40 companies, from three-person startups and AI-natives like Harvey and Decagon to scaled operators like Ramp, Zapier, Notion, and Vercel to public companies like Wix, asking one question: what does...Over the past year I've spent time inside more than 40 companies, from three-person startups and AI-natives like Harvey and Decagon to scaled operators like Ramp, Zapier, Notion, and Vercel to public companies like Wix, asking one question: what does it take for an organization, not just a person, to become AI-pilled?
Nearly every company we visited has individuals who are dramatically accelerated by AI, and almost none have converted that into an institutional advantage. Across the board, token spend is outrunning top-line gains, and cultural buy-in lags technical capability. The companies genuinely ahead aren't the ones that deployed the most tools; they're the ones that redesigned how work flows, how context is shared, and what humans are actually for.
I'll walk through the self-assessment that came out of this research: the questions that reveal how AI-pilled your organization actually is, where the gaps almost always sit, and what it takes to close them.read more
Ann Miura-KoCo-founding Partner, Floodgate
Built for you by(617)-PATRICK, Champion AI speaker
Let AI help you determine which sessions to attend 🤖
The steps below are for copying all the session descriptions and putting them into your own AI model - Claude, Gemini, ChatGPT, Co-Pilot and so on. This is the best way to do it on your laptop or PC. If you don't have your laptop with you, use the phone feature in red at the top of this page.
Tip: This is a large session list. Claude (claude.ai) handles the full file reliably - if you don't have a window open, start one there. Other tools like ChatGPT and Gemini may silently cut off the data - the verification step in the prompt will catch that.
Paste one of these two prompts. The short one is enough; the longer one gives the AI a role, context about the data, and a clear task.
Pick the 3 best sessions for me and say why.
ROLE: Expert conference advisor for The AI Conference 2026 attendees CONTEXT: I'm pasting over 100 sessions for the The AI Conference 2026.
I am an engineer or technical founder shipping AI agents to production. I want the sessions on agent evaluation, inference and security, and good networking with other builders and infrastructure leaders. GOAL: Pick the top 3 sessions for me. OUTPUT FORMAT: List in chronological order. Do NOT number them sequentially. Instead, start each entry with its relevance rank in bold, like: Rank #2 of over 100 - Session Title. Follow with the day and time, then 2-3 sentences on why it fits my goals. VERIFICATION: Before answering, confirm how many sessions you received.
You should have at least 100. If you received significantly fewer,
tell me immediately - my paste was probably truncated.
😎 How cool am I? 😎
Which sessions are right for you?
Describe yourself and what you want from The AI Conference 2026. The more specific you are, the better this works.
You can also say how many you want or when you're free - “give me 10”, “give me 20”, “only Friday afternoon”. You get 3 unless you ask for more.
There is a daily limit on the phone version
so it can stay free for everyone at the event.
However, you can run it as many times as you want using your own AI account at home:
go to the bottom of the page
tap Copy this longer prompt
paste it into Claude, ChatGPT, or Gemini.
Powered by Claude. Your text is sent to Anthropic to generate the ranking, and is logged on this site so the feature can be improved. Please don't type anything confidential.