Inside Google I/O with a DeepMind Exec

Google I/O this year was organized entirely around agents, with Logan Kilpatrick describing the conference as focused on agent-native product development across Google's entire product surface. Jason framed the agentic era as a fundamental shift where useful work happens asynchronously without the user driving every step, contrasting the current moment with AutoGPT from roughly three years prior, which he described as an interesting demo that failed to produce real-world results.

Gemini 2.5 Flash was characterized by Kilpatrick as the best model Google has ever shipped. Originally designed as a smaller, cheaper model for basic chat, Flash has been repositioned as the primary workhorse for the agent era, targeting agentic long-running tasks and coding. Kilpatrick said Flash 2.5 competes with much larger competitor models on coding benchmarks and sits at roughly a Sonnet-level intelligence rather than a mini-model tier. Jason attributed this to distillation, a technique for compressing Pro-level intelligence into a smaller model, crediting Gemini leads Oriol and Jeff with developing the method. Flash 2.5 launched to 900 million users in the Gemini app on day one, which Kilpatrick described as the most widely distributed model launch Google has ever done, though he acknowledged that scale brings its own challenges without detailing them.

Gemini Omni was described as a world model capable of accepting any input type and producing any output type. It fuses VO video generation, Imagen image generation and editing, text-to-speech audio, and the Lyria music model into a single model. The current release is the Flash variant, available in the Gemini app, YouTube, and Flow, with API early access tests planned following Google I/O. Kilpatrick predicted Omni will open a new product category around video remixing and editing, potentially enabling a wave of creator-focused agencies analogous to those that emerged around social media, and suggested effective Omni users could generate millions of followers and views per month.

Gemini 3.5 Pro was in development with multiple iterations running at the time of recording and was announced for availability sometime in the month following Google I/O, not necessarily early in that window. An always-on agentic feature inside the Gemini app was rolling out to trusted testers during the week of Google I/O and was set to reach Gemini Ultra customers the following week.

Managed agents launched in the Gemini API use the same model harness powering the Gemini app's Spark experience. Jason demoed a managed agents use case on stage involving an AI radio show that called approximately seven different models with no orchestration code written, with skills defined in markdown. MCP support for tool calling in managed agents was described as coming soon but not yet available at launch. Jason framed Google I/O as step one of the managed agent story with roughly 50 additional roadmap items remaining.

Antigravity was introduced approximately six months before the recording as an AI-powered IDE and has since expanded into an ecosystem that includes an agent manager for web and desktop, a CLI product, an SDK, and API-level integration via the Gemini API. It serves as the agentic coding layer for production-quality work in large codebases, powers the always-on Gemini Spark consumer experience, is being used internally to build Google products, and supports non-Google models including the OpenAI API. AI Studio is positioned separately for vibe coding, aiming to take users from prompt to deployable product without viewing code, and now supports native Android app building, Google Workspace integration, and app sharing and download to phone. Tens of thousands of people built their first Android app in AI Studio on the first day that feature launched. Native Android support is intended to extend to Android XR glasses, wearables, watches, and Android Auto, though multi-form-factor support was acknowledged as not yet fully functional.

Jason noted that many potential customers do not yet know they need an agentic product, representing an opportunity for builders who can frame the problem correctly, and added that an expectation reset is required every three to six months for people building in the AI space.