If you have used an AI assistant on Android work, you know the ritual. The AI writes the code, you build it, the build fails or the app crashes, and then you copy a wall of red text out of Logcat and paste it back into the chat. The AI never sees your app run. It only sees what you carry back to it.

On October 28, 2026, Google released Android Studio Rabbit 2 as a stable release, and its headline feature removes that middleman. The feature is called Bring Your Own Agent, and it lets more than 40 ACP-compatible coding agents—including Claude Agent, OpenAI's Codex, Junie, and Google's Antigravity—plug into Android Studio and use its tools directly. The question worth asking is not "can AI write my app now," because that was never the bottleneck. It is: what can the agent actually inspect, and which of your debugging chores does that eliminate?

What the agent can see now

According to Google's announcement, agents running inside Android Studio get access to the Android Emulator, the build system, and Logcat. Read that list again, because it maps almost exactly onto the copy-paste ritual. The agent can build the project itself and read the errors. It can boot an emulator, run the app, and watch the logs as it crashes. The loop of write, build, fail, paste, repeat shrinks to write, build, fail, fix.

On top of raw tool access, Android Studio now ships a skills manager with preloaded Android and Firebase skills—guided procedures for things like migrating legacy XML layouts to Jetpack Compose, upgrading to AGP 9, or adapting layouts for foldables. A skill is the difference between an agent that knows Android in general and one that follows Google's own checklist for the specific task.

The built-in agent also gets one specialist worth naming: LeakCanary integration in the Profiler, which explains memory leaks in plain narrative and suggests concrete code changes instead of leaving you alone with a heap dump.

The quieter feature might matter more

Buried in the same release is AI Change Review, and for anyone already letting agents write code, this might be the more important half of the story. Faster code generation just moves the bottleneck to reviewing and verifying. Change Review works on three kinds of changes—agent-made edits, pending commits, and anything in your Git log—and starts by flagging Recommended Fixes, where one agent pass reviews another's work. After that, a Create Walkthrough mode groups the diff into logical themes and walks you through related chunks together, instead of forcing you to read changed files in alphabetical order.

In other words: Rabbit 2 tries to close both ends of the problem. The agent can check its own work against real build and runtime output, and you get tooling to check the agent.

Not an Android Studio user? There is a second door

The same week, Google's developer documentation added official Android CLI plugins for Claude Code, Codex, and Antigravity, for people who work in a terminal or a different editor. Each plugin hands the agent an android-cli skill that can create projects, manage virtual devices, inspect the live UI layout of a running app, run Journeys, and search the official Android Knowledge Base. Antigravity gets a deeper integration with interactive Compose Previews and a device hub inside the agent window; Claude Code and Codex get the plugin through their own plugin commands, a couple of terminal lines each.

The practical meaning: the "agent that can see your app" workflow is no longer tied to the IDE.

What this does not include

A few realities before anyone installs anything. These are integrations, not subscriptions. Whichever agent you pick, you sign in with your own account for that provider, and that provider's pricing, quotas, and regional availability all apply as usual. Rabbit 2 does not come with Claude or Codex bundled in. And while the release is stable and the plugin docs are live, the build-and-verify loop described here is from Google's announcement and documentation—I have not installed and tested it myself, so treat the workflow descriptions as the official story, due for real-world verification.

It is also fair to note what even a fully wired agent cannot judge: whether your app feels right, whether the flow makes sense, whether anyone wants it. Logcat access fixes the "does it run and why not" question. The rest of the review is still yours.

The takeaway

The honest way to read Rabbit 2 is as plumbing, not magic. The interesting shift is not that Android Studio gained another chatbot—it is that the chatbot gained eyes. An agent that can read your build errors and your logs by itself removes the most tedious step in AI-assisted Android work: being the courier between your broken app and the thing trying to fix it. If you try it, the detail worth watching is whether you actually paste fewer stack traces. That, more than any feature list, is the test of what changed.