Android Dev Summit 2026: Android Studio now works with the AI agent of your choice. What should your team change?

Google's message at Android Dev Summit on October 28, 2026 was short: Android tooling is no longer tied to one AI model or one IDE. The most concrete announcement is that Android Studio Rabbit 2 is now stable, and Bring Your Own Agent (BYOA), which arrived in preview in September, is now considered ready for production use.

Rather than walking through every announcement, this post focuses on what an Android team should actually do with them.

What was announced

Android Studio Rabbit 2: bring your own agent

With Rabbit 2 you can pick agents such as Claude Agent, OpenAI Codex, JetBrains Junie or Google Antigravity from Android Studio's agent window. Google says more than 40 agents that implement the Agent Client Protocol (ACP) are supported. You set them up under Settings > Tools > AI > Agents, then sign in or provide an API key.

The important part is that the agent is not just a chatbot editing files. Over ACP, Android Studio hands the agent the project structure, build configuration and platform details, and the agent can drive the emulator, read Logcat, resolve symbols and render Compose Previews. According to the September preview announcement, routine actions can run without prompting while riskier ones pause for your approval. If one agent runs out of quota or underperforms, another can pick up the work.

Other additions in Rabbit 2:

• Agent Skills Manager: install and manage ready-made skills for jobs like migrating XML Views to Compose, upgrading to AGP 9 or building adaptive layouts for foldables.
• Change Review: have AI review agent changes, pending commits or a change from the Git log. Long diffs are presented as a walkthrough grouped by theme instead of alphabetically by file.
• Android XR emulators are now stable.

Android CLI: Android development without the IDE

To me, the more strategic move is the Android CLI. Google is exposing much of its tooling to the terminal, and therefore to agents:

• android preview renders Compose Previews with no emulator or IDE. In Google's internal testing, agents using it finished UI tasks 77% faster with up to 3x fewer tokens.
• android device remote provisions a real device from Google's device lab, which you then use over adb like a local one.
• android signals brings crashes, ANRs and memory issues from Android vitals and Crashlytics into the terminal and the agent's context.
• android analyze scores how well R8 shrinks, optimizes and obfuscates an APK or app bundle, with JSON output for CI.

There are also official plugins for Claude Code and Codex. For Claude Code:

claude plugin marketplace add android/android-cli-plugin
claude plugin install android-cli@android-cli-plugin

Performance and build

• R8 Configuration Analyzer ships with AGP 9.3 and ranks your keep rules by cost (./gradlew :app:analyzeReleaseR8Config).
• Jetpack Keep (alpha) replaces broad keep rules with type-safe annotations.
• Compose 1.13 adds recomposition tracing, showing which functions recomposed and why, all the way back to the state write that triggered it.
• Lightbuild is in early preview: a declarative build defined by a strict YAML schema that agents can read easily. Google recommends it for new or experimental projects and explicitly says to stay on Gradle and AGP for production.

Why it matters

Until now, AI assistants on Android had two problems. General-purpose agents couldn't "see" the build system, emulator or Compose Previews, so they wrote code they couldn't verify. And teams were stuck with whatever model the IDE shipped with.

Google is tackling both at once, not by pushing its own model but by opening its tools through an open protocol (ACP) and a CLI. That choice matters: whichever vendor your company already has an enterprise agreement with, you can use that agent on your Android project, and Google lists this explicitly as a reason to pick an agent.

The second point is the feedback loop. An agent that can render a preview, run tests on a real device and read production crash data is far less likely to hand you code that merely looks right. The 77% and 3x figures come from Google's internal tests, so measure on your own project before treating them as fact. But the direction is clear: agents can now check their own work.

What to do in practice

1. Settle your data policy first. Google states that with Gemini Enterprise, source code isn't retained or used to train models. For third-party agents connected through BYOA, data handling depends on that provider's terms. Agree up front on who connects which agent with which account.
2. Try the CLI in CI. Feeding android analyze JSON output into your pipeline to catch R8 regressions in release bundles can pay off even if you never use an agent.
3. Put Change Review before human review, not in place of it. Themed walkthroughs make large agent-generated diffs easier to read, but the final call should still be yours.
4. Keep Lightbuild out of production. Trying it on a side project makes sense; Google itself doesn't recommend it for main apps yet.
5. Put foldables on your roadmap. Google said Play will look closely next year at how apps run on foldables, with details due later this year. The adaptive layout skill is a reasonable place to start.

Final thoughts

The real message of Android Dev Summit 2026 isn't "more AI" but agent-agnostic tooling. Android Studio is becoming a workspace that many agents can plug into rather than a showcase for one model, and the Android CLI brings the IDE's capabilities to the terminal. If your team already works with Claude Code, Codex or another agent, you can now give it the same context and verification tools on Android projects. In the short term, the easiest wins will probably come from the "boring" parts, like R8 analysis and production signals in the CLI.