The Android Dev Summit ran on 28 and 29 October in Mountain View, and the headline feature of Android Studio "Rabbit 2", now stable, is not a new Gemini capability. It is a settings pane that lets you install somebody else's agent.
Settings > Tools > AI > Agents opens a marketplace of more than 40 coding agents that speak the Agent Client Protocol, an open standard for connecting a model harness to an editor. Google's documentation names Claude Agent and Codex; heise adds JetBrains' Junie and Google's own Antigravity. You can install several and switch between them as you work, and the outside agent inherits everything the IDE knows: symbol resolution, the Gradle build graph, Compose Preview rendering, the emulator, Logcat.
The most telling detail is in Google's own demonstration. To show off memory-leak hunting through the new Android CLI, the team pointed Claude Code — running Claude Opus 5.5 — at a list of critical user journeys with a single prompt: find and fix the memory leaks in this app. Claude spread the journeys across separate devices in parallel, ran the new `android profiler leakcanary` subcommand, followed each reference chain back into the source, patched the code, and re-ran every journey to confirm the leaks were gone. Google published the benchmark chart with a competitor's model on it.
That is a striking concession from the company that ships one of the frontier models these agents compete with. The reasoning looks less sentimental than structural: developers were already running Claude Code and Codex in a terminal beside Android Studio, and the IDE was losing the split. Hosting the agent keeps the window.
The rest of the release is built around that bet, and almost every number Google quotes is about tokens rather than intelligence. Rendering Compose Previews from the terminal with `android preview`, without an emulator, let agents finish UI tasks 77% faster in internal testing while using up to three times fewer tokens. Wiring LeakCanary into `android profiler leakcanary` — analysing heap dumps on the developer's machine instead of the device — made the most capable models roughly twice as fast with up to 25% fewer tokens. `android device remote` provisions a real handset from a Google device lab so an agent can install and test on hardware, and `android signals` pulls production crash and ANR data into the same loop.
Google also shipped the plumbing in the other direction. Android CLI is now an official plugin for Claude Code and for Codex, giving either agent the emulator, live UI inspection, journeys, and the official Android knowledge base; and Change Review adds a `/review` command that reports potential problems in agent-written diffs, walks a human through long changes, and can hand the fixes back to the agent.
Then there is Android Bench 2.0, a leaderboard for models and agent harnesses on tasks designed to take an Android engineer days — migrating a UI, building an app from mocks. Solutions are graded on whether they actually work, not whether they compile, and each is scored for cost and time alongside quality. There are more than 30 official Android skills across 20-odd areas, each backed by evaluations, plus policy and security auditing skills for Play submission. Lightbuild, an early preview of a declarative YAML build, suggests where the toolchain itself is heading.
Two things are worth watching. First, the referee problem: Google now runs the scoreboard on which coding model is worth its tokens, and Google sells a model on that board. A vendor leaderboard is useful evidence, not neutral evidence. Second, the moat has visibly moved. If any agent can drive the emulator and any model can be swapped mid-project, the differentiator is no longer the model — it is the toolchain, the skills, and the depth of project understanding behind them. That is precisely the ground Google has spent fifteen years owning, and it explains why the company is willing to let Claude and Codex stand on it.
