Google I/O 2026: Gemini 3.5 Flash, Antigravity 2.0 and managed agents in the Gemini API
Google I/O 2026 took place on May 19, and the message for developers was very direct: agents everywhere. A new model, a rewritten agent tool, an API that hands you a ready agent in one call, and even a proposed web standard for agents to talk to sites. I spent the last few days reading Google’s summary and the developer keynote recap, and here I organize what seemed most relevant.
Gemini 3.5 Flash
The model of the moment is Gemini 3.5 Flash. Google says it “outperforms Gemini 3.1 Pro across almost all benchmarks” while running four times faster than other frontier models. The post doesn’t include benchmark numbers, so I treat it as a vendor claim until I see independent evaluations.
If the claim holds, the practical consequence is simple: the “Flash” line, normally the cheap and fast option, becomes a candidate for default model in many applications that use a bigger model today. It’s also the model underneath Managed Agents, which says a lot about where Google wants it used.
Antigravity 2.0, CLI and SDK
Antigravity 2.0 became a standalone desktop app, designed as a “home” for agent interaction. It runs multiple agents in parallel, has dynamic subagents and scheduled background tasks, and integrates with Google AI Studio, Android and Firebase. The keynote recap adds cross-platform terminal sandboxing, credential masking and hardened Git policies.
Two important pieces came with it:
- Antigravity CLI: a terminal tool for creating agents without a graphical interface. Google encourages Gemini CLI users to migrate to it.
- Antigravity SDK: programmatic access to the same agent harness Google uses in its own products, optimized for Gemini models and with the option to host custom agents on your own infrastructure.
For enterprises, Google Cloud customers can connect Antigravity directly to their projects. And there’s a new plan, Google AI Ultra at $100 per month, with Antigravity usage limits five times higher than AI Pro.
Managed Agents in the Gemini API
This is the announcement I most want to try. With a single API call, you create an agent that reasons, uses tools and runs code in an isolated Linux environment. It’s built on the Antigravity harness and Gemini 3.5 Flash, and it’s available through the Interactions API and AI Studio. Environments persist across follow-up calls, and you can define custom agents with instructions and skills in Markdown.
In practice, this takes the most tedious part of putting an agent into production off your plate: provisioning a sandbox, keeping state across turns and isolating code execution. The price is the usual one with managed services: less control over the environment and more dependence on a platform.
AI Studio, Android and the web
Google AI Studio gained native calls to Workspace APIs, one-click project export to Antigravity, native Android app generation from a prompt (with Kotlin support), one-click deploy to Cloud Run and Firebase support. On Android, the Android CLI reached stable, letting agents directly use Android Studio’s tooling, such as downloading the SDK and running apps on devices.
On the web, Google described WebMCP as “a proposed open web standard” for exposing structured tools to browser-based AI agents, with an experimental origin trial starting in Chrome 149. It also announced a preview of Modern Web Guidance, with expert-vetted skills for more than 100 use cases, and Chrome DevTools for agents, which lets them run quality audits and emulate user experiences.
My take
I/O made it clear that Google wants to own the entire agent stack: model, harness, runtime environment, IDE, terminal and even the protocol between agent and page. It’s coherent, and Managed Agents solve a real problem. But the push to migrate from Gemini CLI to Antigravity CLI also shows how fast these tools are still changing, and that weighs on deciding where to invest.
What I’d watch closely is WebMCP. If it becomes a real standard, it changes how we think about sites for agents, moving from “the agent reads the DOM and hopes” to “the site declares what the agent can do”. For now it’s a proposal and an origin trial, so it’s worth testing without betting the product on it.