Google's I/O 2026 keynote had the obligatory model-number bump — Gemini 3.5 — but that wasn't the story. The story was a stack of announcements that, read together, describe one strategy: don't try to win the model race on raw capability. Own the rails that AI agents run on.
What Google actually shipped
Almost nothing announced for developers was a chatbot.
- Antigravity 2.0 and a new CLI — Google's agent-first development platform. You can spin up specialized subagents, and it ships with terminal sandboxing, credential masking, and hardened Git policies. A new Antigravity SDK gives programmatic control of the agent harness on your own infrastructure, and Managed Agents in the Gemini API removes the setup entirely: one call, Google hosts it.
- WebMCP — a proposed open web standard that lets a site expose JavaScript functions and HTML forms as structured tools a browser agent can call directly. The experimental origin trial starts in Chrome 149, with Gemini-in-Chrome support to follow.
- Android — a stable Android CLI so agents can drive Android Studio, plus a migration agent that ports React Native, web, or iOS apps to native Kotlin (weeks to hours), and Android Bench, an LLM leaderboard for Android tasks.
- Chrome DevTools that an agent can operate, and an HTML-in-Canvas API.
The strategy underneath
Notice what these have in common: each one puts a Google surface — the browser, the OS, the cloud — between an AI agent and the work it's trying to do.
Antigravity is where you build and run agents. Managed Agents is Google hosting them. WebMCP is the web itself becoming agent-readable, with the standard authored by Google and shipping first in Chrome. The Android CLI and migration agent make agents that build Android apps, which conveniently defends Android.
This is Google playing the only game it has ever been great at: distribution. It doesn't need the single best model if it owns the browser most of the planet uses, the most popular mobile OS, and the cloud the agents call home. The model is a component. The substrate is the moat.
WebMCP is the tell. If agents become how people actually use the web — booking, buying, filling in forms — then the interface between a website and an agent is the next SEO, and whoever defines it sets the terms. Google intends to be the author.
My read
Strategically, this is the right move, and it's more honest than pretending Gemini wins on benchmarks alone. Distribution is Google's hand. It should play it.
Two things should make developers cautious, though. First, "an open standard that ships first in Chrome and works best with Gemini" is the exact shape of AMP — and the web spent years unwinding AMP. A standard that's open in name and Google-shaped in practice is not a gift. Second, there's already a de facto agent-tool protocol in Anthropic's MCP: vendor-neutral and everywhere. WebMCP is Google's browser-flavored answer to it, which means the next year quietly includes a standards fight, and developers hate nothing more than implementing the same thing twice.
So the question isn't whether Gemini 3.5 tops a leaderboard. It's whether the ecosystem adopts a Google-authored browser standard for agents, or whether the neutral protocol that already won stays the lingua franca. Google is betting its distribution can make WebMCP the default. History says the web gets a vote.
Reporting from the Google I/O 2026 developer keynote recap.
