| 9:00 AM | Doors Open | Come in, network, get some coffee and snacks |
| 9:30 AM | Main Theater: Welcome Address | |
| 10:00 AM | Main Theater: Opening Keynote | Opening Keynote by Richard Seroter |
| 11:00 AM | Main Theater: Martin Omander - 7 mistakes I've seen in AI apps | Many apps nowadays include AI, to improve the user interface or to offload tasks from the user. But as with any new technology, it's easy to make mistakes. Martin has seen many of them in the field, and made some himself. In this session he will present the top 7 mistakes that cause poor user experiences or runaway bills in apps that use AI. |
| 11:00 AM | Workshop: Angelina Millare - From Static to Magic: Generative UI with Replit + Antigravity SDK | Discover how agentic development workflows using the Antigravity SDK make it possible to build more adaptive, personalized digital experiences at unprecedented speed using natural language. |
| 12:00 PM | Main Theater: Alec Dibble - Beyond the Chat Window: Four Modes of an AI-Forward Engineering Organization | Giving every engineer an AI assistant is not the same as building an AI-forward organization. AI can change not only how quickly we complete tasks, but how we discover work, make decisions, and collaborate. How do we think about that opportunity beyond individual productivity?
This talk introduces a practical framework for four modes of AI use: Explore, Execute, Observe, and Delegate. By distinguishing between understanding and action, and between on-demand assistance and ongoing responsibility, the framework offers a way to identify AI opportunities and understand how different modes can complement one another.
We’ll explore how these opportunities—and the approaches to implementing them—change at the individual, team, and company levels. What helps one person move faster may not help a team work better, and optimizing a single workflow is different from improving an organization’s overall effectiveness.
Connecting these perspectives opens up a broader conversation about AI adoption: not just accelerating existing tasks, but rethinking how work gets done. |
| 12:30 PM | Courtyard: Lunch | |
| 1:30 PM | Main Theater: Mayuresh Agashe - Beyond device="tpu": Onboard, Tune, & Serve PyTorch Models Natively on Cloud TPU | At PyTorch Conference this year, our team at Google open-sourced two projects: TorchTPU, which makes the TPU a native PyTorch backend, and the TPU agent suite, a harness-agnostic agentic framework that teaches coding assistants like Claude Code and Antigravity to dispatch specialist TPU agents. TorchTPU gets you running with one line: device=""tpu"". This session is about everything after that line.
In this talk, we'll bring a Hugging Face model to Cloud TPU, twice. First by hand: checking the hardware, preflighting the model, checking for code for other accelerators and custom kernels (and what to do when you find them), checking numerical parity, then serving and profiling it. Then with the TPU agent suite, where agents do the same job, show their evidence, and ask before the big moves.
You'll leave with:
A practical checklist for bringing a PyTorch model to a new accelerator
The decode-loop bug that's probably in your code too, and how to fix it
Lessons from building agents that run hour-long engineering jobs: profile before you tune, check before you hand back, and keep a human in the loop |
| 2:00 PM | Main Theater: Stephen Wylie - Throw a "House Party" with Spec-Driven Development Agents | You may have heard of spec-driven development (and if not, we'll briefly review it), but we turn it up to 11 with the idea of the House Party. If you have ever struggled to build a multi-faceted, well-architected system using plain specs and vanilla agents, then consider inviting this set of specialist agents who are already well-versed in many different domains. These agents are constructed from many previous trials and tribulations in their domains, and are ready to learn more and further hone their skills.
If a massive set of agents isn't enough, we also introduce the concept of the ""house"" itself -- a series of rooms that go beyond defining the agent steering (i.e., ""house rules"") at just a project level, but now set the vibe for the way the agents operate, depending on exactly what needs to be accomplished. Now anyone with any kind of stake in the project, from dashboarding to proof-of-concepts to working on security or authorization concerns, can leverage the right set of agents with the right set of rules to more safely present code changes in an appropriately restrained manner.
This talk will explain the concepts of the agents, rooms, and even what gets served (i.e., what strength of model given the task) in-depth and walk you through the steps necessary to set up a house with multiple agents on your own. This way, you can immediately get something positive and productive out of the house party running on top of your code base! |
| 2:00 PM | Workshop: Flavia Ballabene - Building a Multi-Agent System | Single LLM prompts can only take you so far, real-world automation requires specialized agents working together. In this hands-on workshop, you'll learn how to design, coordinate, and deploy a multi-agent system from scratch using Python, Gemini, and Streamlit. Discover how to assign specialized agent roles, pass context between tasks, and leverage the Model Context Protocol (MCP) to seamlessly connect your agents to external tools and data sources. Bring your laptop and leave with practical architecture patterns and a fully working local setup. |
| 3:00 PM | Main Theater: Boris-Wilfried Nyasse - Off the Edge: Grading Google AI Edge and Gemini Against a Practitioner’s Eye | 𝗔𝗜 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗵𝗲𝗹𝗽𝘀 𝘂𝘀 𝗹𝗲𝗮𝗿𝗻, 𝘄𝗿𝗶𝘁𝗲 𝗮𝗻𝗱 𝗱𝗲𝗰𝗶𝗱𝗲. 𝗖𝗮𝗻 𝗶𝘁 𝗵𝗲𝗹𝗽 𝘂𝘀 𝗺𝗼𝘃𝗲 𝗯𝗲𝘁𝘁𝗲𝗿? Getting better at a physical skill takes someone who can see what you cannot: your own body in motion. That asks a machine to understand space rather than words, then say something worth acting on. This talk asks how far that has come, using an application built for the purpose and a sport where the answer can be checked: jumping rope. Someone who practises it can grade the answer, so the model gets no benefit of the doubt. It runs on Google Cloud. Google AI Edge’s Pose Landmarker measures the body. Gemini video understanding is asked for the harder thing: not a description of the video, but advice that changes the next session. How far does a pose model go on real footage, when the object this sport is built around fills 0.058 % of the picture and blurs in every frame? Is Gemini genuinely useful as a coach, or only convincing? And when the measurement and the advice disagree, which one do you believe? We walk the whole route, run the app live on real sessions, and close with what practitioners made of the coaching and a grade for each. |
| 3:00 PM | Workshop: Kiruthika Subramani - Self-Refining Agents That Earn Their Own Upgrades on Google Cloud | What would it take for an AI agent to deserve its next version, instead of just generating one itself? That question is the whole build. As it talks to customers, it takes notes on how to get better and writes them into its own skill files, and for the deterministic tasks behind every reply, the way it classifies, clusters, and recommends, it proves itself against ground-truth labeled data. Leave a thumbs up or a thumbs down and it replays the conversation, learns from it, and drafts a sharper version of itself. But here the testing happens in a way you would not expect. Before any version ships, a friend agent steps in and plays the customer, pushing it with real questions in real time, and only when it proves better than the version it came from does a human give the final yes. This is a self refining, self looping agent running on open-source models auto-scaled on Google Cloud, earning every upgrade by proving it is genuinely better, not just a prompt we talked into looking good. |
| 4:00 PM | Closing and Prizes | |