Google opens Android to third-party AI agents: what changes

Google has made its move in the mobile development arena and, surprisingly, it has done so by opening doors rather than closing them. At the Android Dev Summit held this week, the company announced that third-party coding agents —such as Anthropic's Claude Code or OpenAI's Codex— will be able to deeply access its Android toolchain. We are talking about real devices, performance profilers and, most relevant for any team maintaining an app in production, production crash reports.

The news is significant because Google has not taken the opportunity to push developers toward its own Gemini model. Matthew McCullough, vice president of product for Android, has emphasized the idea of giving developers the freedom to choose their own models and agents. Adarsh Fernando, head of developer tools, sums it up as giving teams and their agents “the tools and knowledge to build high-quality Android apps faster”.

What 'bring your own agent' means

The Bring Your Own Agent feature, now shipping as stable in the Rabbit 2 version of Android Studio, supports more than 40 coding agents that speak the open Agent Client Protocol. Among them are Anthropic's Claude Agent, OpenAI's Codex and Antigravity, Google's own agent. A developer can install several and switch between them without leaving the IDE.

The technical detail that matters: connected agents inherit the IDE's own reading of the project —its structure, build configuration, symbols and tools such as Compose Preview, the emulator or Logcat—. That is, they do not have to rebuild that context through code searches and guesses in the terminal. That context saving is, in practice, what separates a useful agent from one that goes in circles.

Alongside this comes Change Review, a feature that reviews an agent's pending changes for common problems and can guide the developer through a broad set of modifications before applying anything. Google also extends the reach outside the IDE with official Android CLI plugins for Claude Code and Codex, which cover project creation, emulator management, UI inspection and Android documentation.

Previews from the terminal and remote devices

Another of the announcements is the android preview command, which allows agents to render Jetpack Compose previews directly in the terminal, without launching the IDE or the emulator. Google claims that agents using it completed UI tasks 77% faster and consumed up to three times fewer tokens, although it shares few details about how it measured those results. The CLI also adds android device remote, which gives agents access to physical devices in Google's lab for testing.

In the debugging arena, Google gives agents more ways to investigate problems once the app is built: performance, production crashes and memory leaks. In a demonstration, Claude Code worked with LeakCanary, the open source leak detection library, to review critical user paths with a single developer instruction. The agent ran the paths on several devices in parallel, analyzed memory dumps through the CLI integration and traced the leaks to the guilty code. It then applied the fixes and repeated the tests to confirm they worked.

Google maintains that these tools helped capable models complete those investigations almost twice as fast and with up to 25% fewer tokens. A complementary command, android signals, channels data from Android vitals and Firebase Crashlytics into the same terminal flow, which gives agents a direct line to production crashes, ANRs and memory problems.

The nuance that should not be ignored

The demonstration is flashy, but it is worth reading it calmly. The news itself recalls a fact that invites caution: an Anthropic scanner flagged some 29,000 possible vulnerabilities in open source code and only a few hundred were actually fixed. That is, autonomous fault detection can run ahead of the ability to fix them. And none of this removes the human from the loop: some saw Claude Code close a Linux kernel patch and reviewers later catch a race condition.

Google also presents Android Bench 2.0, a test bench to evaluate how models and agents perform on more complex Android tasks, from building an app from mockups to migrating an existing interface. Some tasks are modeled on work that would take an Android engineer several days, and the benchmark also measures completion time and cost.

What we do with this in an SME

If your company has its own Android app or an internal team that maintains it, this announcement opens an interesting door but requires a clear head. Our recommendation, in three steps:

• Evaluate before adopting. Testing an agent on a small project and measuring real time and cost is more sensible than rushing to integrate it into the main workflow. The percentages Google publishes come from its own lab and its own metrics.
• Keep human review. Change Review and automated tests help, but a developer's judgment remains the final filter. As we saw with the kernel episode, an agent can get the patch right and the context wrong.
• Watch token spending. As we mentioned when discussing how Claude Code drives up enterprise spending: the token savings Google promises only materialize if the agent does not spend the day exploring aimlessly.

And a security warning that links to what we have been seeing on the blog: the more autonomy we give an agent over devices, profiles and production data, the more important it is to control what it can and cannot touch. It is the same logic we apply when we review GitHub's npm firewall: automation accelerates, but the team sets the limits.

In short: Google has decided that Android should be an open field for AI agents, and that is good news for freedom of choice. The ball is now in the court of development teams, who will have to decide how much of their workflow they leave in the hands of an agent and how much they reserve for human judgment.

Source: The New Stack. Analysis and adaptation: ForgeNEX.