Google is granting AI coding agents like Anthropic's Claude Code and OpenAI's Codex deeper access to Android development infrastructure, including physical devices, performance profilers, and production crash reports, allowing these systems to hunt down bugs, test modifications, and confirm their own solutions work. The company announced the changes Wednesday at Android Dev Summit, expanding both Android Studio and the terminal-based Android CLI to work with more than 40 third-party agents rather than pushing developers exclusively toward Google's own Gemini model. According to Matthew McCullough, vice president of product for Android, the approach emphasizes letting developers select their preferred models and coding agents.
The company's Bring Your Own Agent feature now works with over 40 coding agents that communicate through the open Agent Client Protocol, including Claude Agent, Codex, and Google's own Antigravity, and the feature is now stable in Wednesday's release of Android Studio Rabbit 2. Developers can install multiple agents and switch between them inside the IDE, with connected agents inheriting the IDE's understanding of project structure, build configuration, symbols, and tools like Compose Preview, the emulator, and Logcat—context they'd otherwise reconstruct through code searches and terminal commands. A new Change Review feature scans pending agent modifications for frequent problems and walks developers through extensive change sets before application, while official Android CLI plugins for Claude Code and Codex handle project creation, emulator management, UI inspection, and Android documentation. A newly introduced android preview command allows coding agents to render Jetpack Compose previews straight from the terminal without opening an IDE or emulator, and the company reports agents using it finished UI tasks 77% faster while consuming up to three times fewer tokens, though it disclosed little about measurement methodology. The CLI also adds android device remote, granting agents access to physical devices in Google's lab for testing purposes.
Google is also expanding agent capabilities for investigating post-build problems, including performance bottlenecks, production crashes, and memory leaks. In one demonstration, Claude Code collaborated with LeakCanary, an open source leak-detection library, to examine critical user journeys for memory leaks using a single developer instruction—running the journeys across multiple devices simultaneously, analyzing heap dumps through the CLI's LeakCanary integration, tracing leaks to problematic code, applying fixes, and rerunning tests to verify success. The report states the tooling helped capable models complete these investigations nearly twice as fast while using up to 25% fewer tokens. A companion android signals command now feeds Android vitals and Firebase Crashlytics data into the same terminal workflow, providing agents direct access to production crashes, ANRs, and memory issues.
Adarsh Fernando, senior product manager for Android developer tools, describes the goal as equipping developers and their agents "with the tools and knowledge to build high-quality Android apps faster." The report notes that none of this eliminates the human in the loop, pointing to developers who watched Claude Code land a Linux kernel fix only to have reviewers catch a race condition, and observes that autonomous bug-finding can outpace autonomous bug-fixing—an Anthropic scanner flagged roughly 29,000 possible vulnerabilities in open-source code while only a few hundred were ever repaired. Google is also launching Android Bench 2.0 to evaluate how models and agent harnesses perform on more complex Android development tasks, testing whether completed apps actually run with assignments ranging from building an app from mockups to migrating an existing interface, with some tasks modeled on work that would occupy an Android engineer for several days, and the benchmark also measures completion time and cost. The platform-agnostic approach signals that Google is betting developer loyalty depends more on tool flexibility than model exclusivity, even as it positions Android as the testing ground where any AI agent can prove its worth. If the economics hold—faster completion with fewer tokens translates to lower costs—the pressure shifts to human reviewers who must verify that speed hasn't traded away correctness.