Android Studio Rabbit 2 Opens to 40+ AI Coding Agents

Android Studio Rabbit 2 went stable at Android Dev Summit 2026 with "Bring Your Own Agent": more than 40 coding agents now run inside the IDE.

Developers can install Anthropic's Claude Agent, OpenAI's Codex and Google's own Antigravity from a built-in marketplace, alongside 40+ agents that speak the open Agent Client Protocol.

What actually shipped

The agents get real IDE access rather than a chat box: project structure, symbol resolution, the emulator and Logcat. Two pieces matter most in a daily loop:

• Compose previews without the emulator. Agents can render Jetpack Compose previews without running an Android Emulator or deploying an APK, and Firebase-powered device streaming lets developers test on real devices remotely.
• Android CLI as an official plugin. It is now an official agent plugin for Claude Code, Codex and Google Antigravity, covering device testing, UI validation, performance profiling and app quality analysis — plus LeakCanary for memory leaks, Perfetto traces for cold startup and UI jank, and an Analyze feature that checks bundles against the same logic Google Play uses, before upload.

Google also shipped more than 30 Android Skills across more than 20 areas, feeding current API documentation to agents for Play policy checks, permission security, Compose performance and foldable layouts.

The speed numbers are Google's own

Two figures are doing the marketing work here, and both come from Google's internal testing. Google says rendering Compose previews in the terminal made agents finish UI tasks 77 percent faster with up to three times fewer tokens. In a Google A/B test with Claude Opus 5.5, the company says the CLI tool found and fixed memory leaks nearly twice as fast with about 25 percent fewer tokens, because heap dumps are evaluated on the computer rather than slowing down the device.

The source reporting is explicit that no independent benchmark yet confirms Google's speed, token or leak-fix claims on real-world Android codebases. Treat them as a reason to run a trial, not a number to put in a sprint estimate.

What it costs you

The announcement does not state pricing changes for the agents themselves — each agent still runs on its own provider's terms. So the IDE is free to open; your token bill is unchanged and still per-vendor.

That makes this a bake-off, not a feature comparison. Protocol compatibility is not quality: 40+ agents means 40 agents that connect. Run one real task you already know the answer to — a bug fix with a failing test, a permission audit — through two agents and compare the diff, the test result and the token bill.

The in-app agent APIs are still early

Beyond the IDE, Google is shipping APIs that turn apps into agent tools, and the stage labels matter:

API	Stage	What it does
ML Kit Agentic API	Alpha	Extends the ML Kit Prompt API with Gemini Nano 4 for on-device agents; executes functions, saves history, manages the orchestration loop, adds automatic guardrails
AppFunctions	Beta	Android's unified tool-calling layer, exposing app functionality to on-device or system agents through Android MCP
A2UI	Alpha	Lets agents return interactive interfaces built from an app's own Compose components instead of plain text
Connected Apps Platform	Preview	Submits your app into Gemini and Search; beta for U.S. developers via waitlist

On the distribution side, developers integrate with a cloud-hosted MCP server or Android MCP using the AppFunctions APIs, then register, test and submit through the Google Connected Console, choosing Search, the Gemini app, or both. Google already lists over 50 Connected Apps in Gemini, including Adobe, Experian, Localiza and Peloton, and says those apps reach Search over the next few weeks.

What to watch next

No general-availability dates were given for the alpha and beta APIs, and Google did not say when the Connected Apps preview opens beyond U.S. developers. Nor has it said whether the IDE marketplace will rank, recommend or restrict agents — the detail that decides whether "40+" stays a real choice. Also worth tracking: Android Bench 2.0, which measures models on multi-day projects by functionality, time and cost.

Prototype on the alphas if you want, but ship the path that works without an agent first.