Android Studio gets "Bring Your Own Agent": Run 40+ coding agents via an open protocol

Android Studio gets "Bring Your Own Agent": Run 40+ coding agents via an open protocol

With "Bring Your Own Agent" in Android Studio Rabbit 2, developers no longer have to use Google's own AI tools and can instead run more than 40 coding agents – including Claude Agent, OpenAI Codex and Google Antigravity – directly in the IDE. The company is also launching new APIs designed to bring agentic features locally onto the phone itself.

The news

At Android Dev Summit, Google announced a series of updates that put AI agents at the center of Android development. The most prominent is "Bring Your Own Agent" (BYOA) in Android Studio Rabbit 2: developers are no longer limited to Google's own AI tools, but can use the open Agent Client Protocol (ACP) to run more than 40 different coding agents directly in the IDE. That's according to The Decoder (Maximilian Schreiner), which reported on the announcements on 28 October 2026 (The Decoder).

This is an opening that breaks from the pattern where development tools typically lock users to the vendor's own models. With an open protocol layer, the choice of agent becomes a developer decision rather than a platform decision – at least in principle.

How it works in practice

Via a built-in marketplace, developers can install and run Anthropic's Claude Agent, OpenAI Codex and Google Antigravity – and they can run simultaneously. Agents connected via ACP get direct access to key IDE tools: the project structure, symbol lookup, the emulator and the Logcat system.

That is where the mechanism lies: a coding agent that only sees the source code is considerably weaker than one that can build the project, launch the emulator, read log output and navigate symbol references the way a human developer does. By exposing these tools through a common protocol, Android Studio becomes a place where multiple agents can work side by side with the same level of access.

Agentic features on the phone itself

The other main part of the announcements concerns mobile devices. With new interfaces – the Gemini Nano 4-powered Agentic API and the Model Context Protocol – developers can run agentic features locally on mobile devices and connect app functionality directly to Google Search and the Gemini app.

This opens up a scenario where an app's services become available to the agent ecosystem outside the app: an action that previously required the user to open the app could, in principle, be carried out via an agent accessing the app through these protocols. Local execution via Gemini Nano also means that some agentic functionality does not require a network connection.

Google's own performance figures – not independently verified

Google presented several performance figures for the new tools. All come from the company's own tests and are not independently verified:

• 77% faster UI tasks and up to three times fewer tokens: According to Google's internal tests, agents complete UI tasks 77% faster and use up to three times fewer tokens with the new terminal-based Compose preview rendering.
• Nearly twice as fast memory-leak debugging: In a Google A/B test with the Anthropic Claude Opus 5.5 model, the CLI tool with integrated LeakCanary helped find and fix memory leaks almost twice as fast, with roughly 25% reduced token use.

These figures are concrete signals of what Google believes the tools achieve, but they are the company's own measurements, reported via the secondary source The Decoder. Independent confirmation is not available in the source material.

The rest of the package

Google also launched a series of surrounding tools:

• Android Bench 2.0: An updated version of the company's leaderboard for language models, which now measures models on multi-day development projects – assessed on functionality, time spent and cost.
• Change Review and a Jetpack Compose migration assistant: Tools for code review and the transition to Compose.
• Lightbuild: A declarative build system.
• XREAL AURA: In collaboration with XREAL, Google launched a pair of smart glasses costing $1,099 that run Android XR, display virtual workspaces and use Gemini to analyse the physical environment around the user.

What remains to be answered

The story is based on secondary reporting from The Decoder; Google's own blog posts and documentation are not part of the available source material. Concretely, that means:

• The performance tests are thus independently unverified, and the test setups (which tasks, how many agent runs, which baselines) have not been published in the sources.
• The details of several of the announcements – including the scope of the new mobile-side APIs – are currently known only in broad outline.
• The XREAL AURA pricing and specifications are single-source details provided by Google and XREAL via the same source.

The concrete substance – more than 40 agents, simultaneous execution, an open protocol, and new local APIs – is, however, consistently and coherently reported, and forms the core of what Google is now trying to do with Android development: to make the agent ecosystem part of the platform itself, both on the development side and on the phone.