Google spent two weeks of May on stage. The Android Show: I/O Edition on May 12 previewed Android 17, and Google I/O on May 19–20 followed with a new Gemini model family and a long list of developer tools. Read together, the message was hard to miss: Google expects software to be built — and increasingly used — by AI agents.
The Android Show: Android 17
The Android Show introduced Gemini Intelligence, a set of agentic features across phones, watches, cars, glasses and Google's new Googlebook laptops, with users confirming sensitive actions. Android 17 also brings Rambler, a dictation feature that aims to drop filler words and handle switching languages mid-sentence. Two announcements reach beyond Android itself: a wireless iPhone-to-Android transfer built with Apple, and Quick Share working with AirDrop on phones from more manufacturers later this year.
For developers there was a small but welcome privacy change: Android 17 lets an app request one specific contact instead of access to the whole address book.
Google I/O: Tools for Agents
The developer keynote introduced the Gemini 3.5 series of models, and nearly everything around it was about agents:
- Managed Agents in the Gemini API provision an agent with a remote sandbox in a single call.
- Antigravity 2.0, Google's agent-first development platform, gains a CLI, subagents for complex work, terminal sandboxing and credential masking, and an SDK for running the agent harness on your own infrastructure.
- Android tooling for agents: a stable Android CLI that lets AI agents use Android Studio's tooling, open-sourced Android “skills” for best practices, and a preview migration agent that moves apps from React Native, web frameworks or iOS to native Kotlin.
- WebMCP, a proposed web standard for exposing structured tools to browser-based agents, with an origin trial starting in Chrome 149.
What Stood Out to Us
Agents are becoming a platform audience. WebMCP is the clearest sign. Websites have long had two audiences, people and search crawlers; a standard way to offer tools to agents adds a third. It is early, but teams building web products should keep an eye on it.
The tooling assumes agents will write code. An Android CLI built for agents, curated skills and a migration agent are not features for humans clicking through an IDE. They are infrastructure for letting a model operate the toolchain safely — sandboxing and credential masking were announced alongside, not later.
Cross-platform friction is shrinking. An iPhone transfer built together with Apple and Quick Share meeting AirDrop make switching between platforms easier for users. For app makers, that raises the bar for consistency: people will notice when the same app behaves differently on their new phone.
What We'd Watch Next
- How the Gemini 3.5 models perform on real workloads, outside Google's demos.
- Whether WebMCP gathers support beyond Chrome.
- How well migration agents handle real apps, where most of the work is in edge cases rather than screens.