Google Opens Android Studio to AI Coding Agents, Letting Claude and Codex Build Beyond Code

Google is granting AI coding agents like Anthropic's Claude Code and OpenAI's Codex expanded capabilities within its Android development ecosystem. The agents now gain access to performance profilers, connected devices, and production crash reports, allowing them to identify bugs, run tests, and validate their own corrections.

The initiative encompasses both Android Studio and the command-line Android CLI, with Google deliberately choosing to open these tools to third-party agents rather than pushing developers toward its own Gemini model. Matthew McCullough, vice president of product for Android, has stressed the importance of letting developers select their preferred models and agents.

Adarsh Fernando, senior product manager for Android developer tools, outlined the objective in a Wednesday announcement: giving developers and their agents "the tools and knowledge to build high-quality Android apps faster." The announcements took place at Android Dev Summit.

Bring your own agent

Google's Bring Your Own Agent capability now accommodates over 40 coding agents that implement the open Agent Client Protocol, including Anthropic's Claude Agent, OpenAI's Codex, and Google's Antigravity. The feature reached stable status with Wednesday's release of Android Studio Rabbit 2.

Developers can deploy multiple agents and toggle between them within the IDE. Connected agents automatically access the IDE's understanding of a project—its layout, build settings, symbol definitions, and utilities including Compose Preview, the emulator, and Logcat—eliminating the need for agents to reconstruct this context through manual code inspection and terminal commands.

A new Change Review feature examines modifications proposed by an agent for typical problems and can guide a developer through extensive changes before implementation. Google is also extending this capability beyond the IDE through official Android CLI plugins for Claude Code and Codex, which handle project initialization, emulator operation, UI analysis, and access to Android documentation.

Compose previews from the terminal

A fresh android preview command permits coding agents to display Jetpack Compose previews directly in the terminal, bypassing the need to launch the IDE or emulator.

According to Google, agents leveraging this capability finished UI assignments 77% faster while consuming up to three times fewer tokens, though the company provided minimal detail regarding measurement methodology. The CLI also introduces android device remote, granting agents access to physical devices housed in Google's testing laboratory.

Finding memory leaks with Claude

Google is expanding agent capabilities for diagnosing post-deployment problems, including performance degradation, production failures, and memory leaks. In a demonstration, Claude Code collaborated with LeakCanary, an open-source leak-detection tool, to scan critical user workflows for memory leaks based on a single developer instruction.

Claude Code executed the workflows simultaneously across multiple devices, examined heap dumps via the CLI's LeakCanary integration, and pinpointed leaks to specific code sections. The agent then implemented corrections and reran validation to confirm success. Google reports the tooling enabled capable models to finish these investigations roughly twice as quickly while consuming up to 25% fewer tokens.

The demonstration showcases impressive capability, yet it also reflects a cautionary reality: autonomous vulnerability detection can surpass autonomous vulnerability remediation. An Anthropic scanner identified approximately 29,000 potential vulnerabilities in open-source projects, yet only several hundred were ever patched.

A new android signals command directs Android vitals and Firebase Crashlytics information into the same terminal interface, providing agents direct visibility into production failures, ANRs, and memory issues. However, human oversight remains essential—developers who observed Claude Code resolve a Linux kernel issue only to have reviewers discover a race condition understand this limitation well.

Benchmarking agents with Android Bench

Google unveiled Android Bench 2.0, a new evaluation framework for assessing how models and agent systems perform on intricate Android development assignments. The benchmark evaluates whether completed applications function correctly, with tasks spanning from constructing an app based on design mockups to refactoring existing interfaces. Certain tasks mirror work requiring several days from an Android engineer, and the benchmark also tracks execution duration and expenditure.