Google Opens Android Studio to 40+ Outside AI Agents as Bring Your Own Agent Goes Stable

At its Android Dev Summit on October 28, Google stopped insisting that the assistant inside its IDE be its own. Android Studio Rabbit 2, now stable, ships with Bring Your Own Agent: through the open Agent Client Protocol (ACP), more than 40 outside coding agents can run directly in the IDE, installed from a built-in marketplace and switchable per task. Google named Claude Agent, OpenAI's Codex, Junie and its own Antigravity as launch integrations.

The pitch addresses a workflow developers have been stitching together themselves: running Claude Code or Codex in a terminal window beside the IDE just to reach the emulator, Logcat and project context. Agents installed through BYOA get Android Studio's own machinery instead — project structure understanding, symbol resolution, Jetpack Compose preview rendering, the emulator and logs. A new Change Review tool audits agent-generated diffs, generates reports of potential issues, and can be invoked with a /review command for larger changesets.

The quieter half of the announcement may matter more. Google's Android CLI now installs as an official plugin for Claude Code, Codex and Antigravity, and it works without Android Studio open at all — in any terminal, harness or CI pipeline. New commands let an agent render Compose previews without booting an emulator. In Google's internal tests, agents finished UI tasks 77 percent faster using up to three times fewer tokens; those are company-reported numbers, not independent measurements. Google also wired the LeakCanary memory-leak library into the CLI and Profiler, and says an A/B test with Claude Opus 5.5 found and fixed memory leaks nearly twice as fast while cutting token usage by about 25 percent.

Google is also feeding agents documentation explicitly. More than 30 Android Skills ship to cover tasks models are often stale on, from checking Google Play policy before submission to permission handling and foldable layouts. Android Bench 2.0, a leaderboard that scores model-and-agent combinations on multi-day real Android builds by functionality, time and cost, gets a refresh alongside a Lightbuild experiment — a YAML-schema build system both humans and agents can read and edit — and a Canary preview that auto-migrates old XML layouts to Compose.

On the device side, an alpha Agentic API backed by Gemini Nano 4 lands in ML Kit, handling function execution, chat history, the orchestration loop and safety filters entirely on-device. AppFunctions enters beta using the Model Context Protocol as a shared tool-calling layer, and an alpha Agent-to-UI (A2UI) interface lets agents render interactive interfaces from an app's own Compose components rather than plain text. A preview Connected Apps Platform manages app-to-Gemini connections from one console; Google says over 50 partner integrations from companies including Adobe, Peloton and Experian are already live in the Gemini app and will reach web search in the coming weeks.

Strategically, Google has done the arithmetic: the best coding agent changes monthly, so owning the protocol beats owning the agent. ACP plays the same role for coding agents that MCP played for tools, and Android Studio is now an agent shell that treats model vendors as interchangeable parts. The more consequential bet is on-device: Gemini Nano 4 agentic APIs plus MCP-backed AppFunctions position Android apps as things agents invoke, not just things people tap. The caveats are the usual ones — performance claims are Google's own tests, and 40-plus agents at launch means quality will be uneven. But the direction is clear: the next app platform fight is for agent access, and Google just opened its gates first.