Android Studio Rabbit 2 Opens the IDE to 40+ AI Agents, Drops Gemini-Only Lock-In
This releases Android development from vendor lock-in: you can keep your familiar IDE workflow while swapping between Claude, Codex, or Google's Antigravity, and the skills packs push Android-specific best practices into whatever agent you choose. The LeakCanary integration also shortens the debug loop, turning a manual memory-leak hunt into a one-click AI fix.
Android Studio Rabbit 2, released stable on October 28, breaks the Gemini-only constraint of Rabbit 1. The IDE now supports bring-your-own-agent (BYOA): over 40 agents that implement Google's Agent Communication Protocol (ACP) can be selected and hot-swapped in the Agent Panel. Officially named are Claude Agent, OpenAI Codex, JetBrains Junie, and Google Antigravity. The point is no longer which model the tool bundles, but which agent best fits your workflow.
The agents get deep IDE access: they can read project structure, trigger builds, resolve symbols, render Compose previews, operate the emulator, and read Logcat. A new Skills Manager adds 30+ Google-authored skill packs covering migration, AGP upgrades, adaptive layouts, CameraX, performance, security, and Play policy compliance, invoked with @ in chat. Rabbit 2 also adds AI change review with a Guided Walkthrough for long diffs, and integrates LeakCanary into the profiler with on-machine heap analysis, roughly 2x faster analysis and up to 25% lower token usage, plus a "Fix with AI" button.
The article's framing is straightforward: Rabbit 2 turns Android Studio from an AI tool into an agent platform. Agent capability is limited by tool access, and the skills packs inject Android best practices directly into agent instructions. For developers, choosing an agent means matching workflow needs: Gemini Enterprise for compliance, Claude/Codex for raw code generation, Junie for JetBrains users, Antigravity for Android-only work, with Antigravity suggested as a free starting point.
The shift to multi-agent support signals that Android Studio is competing on integration breadth rather than on model quality: the IDE's value is now the toolbox and the ecosystem around the agent, not the strength of any single model.
Giving agents direct access to build tools, emulators, and Logcat changes the failure mode of AI coding: instead of hallucinating from pasted snippets, agents can verify their own output against actual build and runtime state inside the IDE loop.
The Skills Manager is a tacit admission that raw model intelligence isn't enough for Android development; best practices need to be packaged as executable instruction sets, which suggests domain-specific guidance is becoming the key differentiator in AI-assisted tooling.
The LeakCanary integration with 'Fix with AI' is the most concrete sign of the editorial stance that the toolchain loop (write, review, fix) can be closed inside the IDE, reducing the human to reviewing rather than debugging from scratch.
Will a relay/proxy service key work?