Google has made Android Studio Rabbit 2 stable, opening its official Android development environment to more than 40 outside coding agents, including Anthropic’s Claude Agent, OpenAI’s Codex and JetBrains’ Junie, alongside its own Antigravity tool. Announced at Android Dev Summit, the change means developers no longer have to choose between their preferred AI assistant and Google’s IDE, because agents installed through the new Bring Your Own Agent feature gain access to the project’s structure, emulator, logs and build tools.
What developers actually get
Bring Your Own Agent works through a marketplace inside the IDE. Developers open Settings, then Tools, then AI, then Agents, browse the available agents, install more than one, and switch between them. Each agent keeps its own strengths, so a developer can use one for drafting and another for debugging.
The agents connect through the open Agent Client Protocol, a standard for linking code editors and coding agents. That shared plumbing is what lets one integration cover so many tools. Once installed, an agent can see the project’s structure and build setup, resolve symbols, render Compose previews, and drive the emulator and Logcat, rather than only editing files.
Google’s own Antigravity tool sits in the same picker, with access to the latest Gemini models as they launch. The company presents the choice as a matter of fit: a developer might pick one agent for its cost, another for a particular strength, or a third because an enterprise agreement covers it.
Beyond the IDE: Android CLI
The same push extends to the terminal. Android CLI, Google’s tool-first interface for agents, is now an official agent plugin for Claude Code and Codex, so agents working outside the IDE can create Android projects, manage emulators, inspect interfaces, run test journeys and search Google’s Android knowledge base. It works in continuous integration pipelines too, where no IDE is open.
New CLI commands cover the parts of testing that used to need a person at a keyboard. A preview command renders Jetpack Compose previews without an emulator, so an agent can check its own interface work across screen sizes as it iterates.
Device streaming provisions a real phone from a Google device lab, and a device hub view lets developers watch local and remote devices without opening the IDE. Both features aim at the same bottleneck: an agent that cannot see or run what it built hands back code it cannot verify.
Google’s internal testing puts numbers on those gains: agents using terminal previews finished interface tasks 77 percent faster while using up to three times fewer tokens. Those are the company’s own measurements, not independent benchmarks, and they will matter less than how the tools behave on real projects.
The same Dev Summit announcements also opened Google Search and Gemini to outside apps, a parallel move toward letting agents work inside Google’s surfaces rather than only beside them, as covered in Google’s Connected Apps opening. It is the same logic applied to search: meet developers and their agents where they already work.
Checking the agent’s work
Reviewing agent-written code is becoming the bottleneck, and Google has built a response into Rabbit 2. Change Review starts from pending agent changes, pending commits or the Git log, flags recommended fixes, and can walk a developer through changes grouped by theme instead of file by file. It can be invoked with the review command from the agent window.
A skills manager gives developers finer control over the specialised workflows an agent can use, from migrating old XML layouts to Jetpack Compose to upgrading build tooling and adding camera or Firebase features. That control matters because an agent is only as safe as the workflows it is allowed to run.
Android Bench 2.0, a leaderboard that grades models on multi-day Android tasks by whether the finished app actually works, offers a way to compare agents before committing to one. Emulators for XR glasses are now stable too. The benchmark grades on working results, not on code that merely compiles.
Why It Matters
The significance is competitive as much as technical. For years, the maker of a development platform could steer developers toward its own AI assistant simply by making it the convenient default. That default shaped which tools teams learned, bought and built around.
By handing rival agents the same project knowledge, emulator access and logs, Google removes that advantage and forces agents to compete on quality, price and trust instead of distribution. Developers gain leverage, and smaller agent makers gain a route into Android work.
That openness also carries a risk Google does not hide: an agent with deep project access can make large, hard-to-review changes, and its performance claims still rest on vendor testing. Review tools help, but they do not remove the need for a human to check consequential changes.
Enterprise teams get some guardrails through Gemini Enterprise support, with promises that code is not retained for training, plus audit logging and admin policies over approved agents. Those controls will be tested by teams with strict compliance rules. Admins can define which agents and model endpoints a whole organisation may use.
The same agent-first shift is playing out across other tools, as seen in GitHub Copilot’s cross-repository projects, so Android Studio’s move is part of a wider contest over where developers spend their working day. Whoever owns that working day owns the next decade of developer tooling.
Background
Google opened Android Studio to any AI model last year and previewed Bring Your Own Agent in the Canary release in September, before moving it to the stable channel with Rabbit 2. The stable release also brings a LeakCanary update, a new agent skills manager and AI code review, alongside the agent integrations.
Source: Android Developers Blog announcement and Dev Summit tools roundup
0 Comments