Android Studio Rabbit 2 is stable: Bring Your Own Agent, skills and Android CLI

On October 28, 2026, at Android Dev Summit, Google released Android Studio Rabbit 2 to the stable channel. The announcement on the Android Developers Blog says it is ready for production use, and the release notes list it as version 2026.2.2.

We build our apps with Kotlin Multiplatform and Compose and work with coding agents every day, so we read the announcements for one question: what changes for a small team. The facts below come from four Google posts and the release notes. Where we interpret, we say so. We have not run Rabbit 2 on our projects, so this is a reading of announcements, not a test report.

What Rabbit 2 stable includes

Google names these items for the stable release: Bring Your Own Agent, a skills manager, AI code review (Change Review), stable Android XR emulators for glasses, and a LeakCanary integration. The release notes add that Rabbit 2 works with Android Gradle plugin versions 7.1 to 9.5. Not everything announced that day is stable, so we split it up.

Bring Your Own Agent in practice

BYOA connects coding agents through the open Agent Client Protocol (ACP). Google says you can choose from over 40 agents. The blog names Claude Agent, Codex from OpenAI, Junie and Google Antigravity, and the release notes list Claude Agent, Codex and Antigravity as the defaults in the agent window.

To start, open the agent window and select an agent, or pick one under Settings > Tools > AI > Agents, authenticate and send a first prompt. That page is also a registry: you can install several agents and switch between them.

What the agent gets from the IDE matters as much as which agent you pick. According to Google, it can use the Android Emulator, the build system, Logcat, symbol resolution, Compose Preview rendering, terminal and shell execution, the project graph and build diagnostics. For sign-in, the release notes mention personal API keys, consumer subscriptions and enterprise subscriptions, and say the built-in agent stays available at no extra charge.

Google gives three reasons for choosing your own agent: you prefer a specific ACP agent, you want the cost profile of a particular one, or you must use a provider covered by an enterprise agreement. For teams with strict requirements, Google points to Gemini Enterprise, available since Rabbit 1: no retention of code and context, audit logs of agent activity, and admin policies for approved skills, agents and model endpoints.

Skills, Change Review, LeakCanary and XR emulators

Skills manager

Skills are packaged instructions that give an agent knowledge about a task. Android Studio now lets you discover, install and manage them. Open the Skills tab from the agent panel menu, type @ in the chat box and choose Manage skills, or use Settings > Tools > AI > Skills Manager. Studio ships with Android and Firebase skills; the examples named are migrating XML Views to Compose, upgrading to AGP 9, adaptive layouts for foldables and tablets, and integrating CameraX and Firebase services.

Change Review

Google frames review as the new bottleneck: agents write fast, people still verify. Change Review starts a session from the agent changes, pending commits or the Git log, or from the agent chat with the /review command. The agent first lists recommended fixes. Then Create Walkthrough groups related diff chunks into themes instead of alphabetical files.

LeakCanary and XR

The LeakCanary integration lets the agent explain the cause of a leak in plain language and suggest a code change, from the LeakCanary task in the Profiler. The XR emulators for XR glasses, audio glasses and display glasses are stable and run embedded in the Running Devices window. The display glasses emulator accepts your own 360-degree photos, still images or a live camera feed. For audio and display glasses you also need a host phone emulator on a Canary system image.

Android CLI: the same tools outside the IDE

Android CLI is a separate tool for any terminal, any agent harness and CI. The posts list these capabilities:

• ✦Device streaming. android device remote provisions a real device from a Google device lab, powered by Firebase, and you install and test as with a local device. android device hub shows local and remote devices without the IDE.
• ✦Compose previews. android preview renders Compose Previews without an emulator or an APK, so an agent can check its own UI changes across screen sizes.
• ✦Profiling. android profiler captures Perfetto system traces to debug cold startup latency, UI jank and background wakeups, and pairs with an android-profiler skill.
• ✦Analyze. android analyze scores how well R8 shrinks, optimizes and obfuscates an APK or app bundle, with JSON output for CI. For a library it shows size, cost and overly broad keep rules.
• ✦Skills and docs. android skills list and android skills add install skills for any agent; android docs search queries the Android Knowledge Base.

There are official plugins that install the CLI into Claude Code and Codex. These are the commands from Google's post:

# Claude Code, from the terminal
claude plugin marketplace add android/android-cli-plugin
claude plugin install android-cli@android-cli-plugin

# Codex, from the terminal
codex plugin marketplace add android/android-cli-plugin
codex plugin add android-cli@android-cli-plugin

What the Dev Summit posts add

The tools and performance post is a round-up, and several items are early:

• ✦Android Bench 2.0, a leaderboard of models and agent harnesses on long, real Android tasks.
• ✦A Migration Assistant for XML Views to Compose and Lightbuild, a declarative YAML-based build. Both are in Canary, and Google recommends Lightbuild only for new or experimental projects.
• ✦A new R8 Configuration Analyzer in AGP 9.3, run with ./gradlew :app:analyzeReleaseR8Config, and Jetpack Keep in alpha.
• ✦Recomposition tracing in Compose 1.13, and androidx.tracing 2.0 for richer traces.
• ✦Skills for Google Play policy and for permissions and intent security audits.
• ✦Crashlytics heap dumps from real out-of-memory crashes, in early access.

A separate post from the same day adds that the Compose adaptive skill can audit UI for keyboard shortcuts, pointer cursors and hover states. Google also says Play will look closely at how apps run on foldables next year. On November 2, Google published over 50 session recordings with transcripts, in four tracks.

Our reading for a small KMP and Compose team

Everything in this section is our interpretation, not Google's statement.

Workflow

There are now two entry points: an agent inside the IDE, and an agent in the terminal with the CLI plugin. We already use agent-driven development, as we describe in our post on building with AI agents and TDD. For us the CLI is the more interesting half, because it brings devices and previews to an agent that does not live in Android Studio. The skills look Android-first, and the sources say nothing about shared Kotlin code or iOS.

Cost, data and contracts

BYOA separates the IDE from the model bill. If you already pay for an agent, you can in principle use it here, with the built-in agent as a fallback. The flip side is that the context Studio hands over, such as the project graph and build diagnostics, goes to whichever provider you choose. The privacy promises Google describes for Gemini Enterprise do not automatically apply to another provider, so read that provider's terms and your enterprise agreement before connecting it.

Before an agent touches the project

• ✦Permissions. The agent can run shell commands and control emulators. Decide what it may run without asking, and keep signing keys, keystores and local property files out of its reach.
• ✦Skills. A skill is an instruction your agent will follow. Install only what you read, and keep the list short.
• ✦Tests first. Agents are only as safe as the tests that catch them. Have the tests before the change, not after.
• ✦Review. Work on a branch and use Change Review as an extra reader, not as a replacement for yours.

What we do not know

• ✦How Rabbit 2 behaves on our KMP projects, and whether project-aware tools understand common source sets.
• ✦Whether all 40-plus agents work equally well; the sources only say they are compatible.
• ✦What third-party agents cost for real work, and what their providers do with the context.
• ✦How the LeakCanary feature is split between stable and Canary.

What to do

• ✦Update one machine to Rabbit 2 first and check that your AGP version is within 7.1 to 9.5.
• ✦Open Settings > Tools > AI > Agents, connect the agent you already use and run one small task on a throwaway branch.
• ✦Read the privacy and retention terms of that agent's provider before giving it real project context.
• ✦Review the installed skills in the skills manager and remove what you do not need.
• ✦Write down what the agent may run and which files it must never read.
• ✦Try Change Review on a real diff and compare it with your own review.
• ✦If you use agents in the terminal, try the Android CLI plugin and android preview on a Compose screen.
• ✦Run android analyze on a release bundle and keep the result as a baseline.
• ✦Treat Canary items and vendor speed numbers as unverified until you test them.

Sources

• ✦Your favorite agents are now integrated in Android Studio (Android Developers Blog, October 28, 2026)
• ✦Agentic and Flexible: A next era of Android development (Android Developers Blog, October 28, 2026)
• ✦What's new in Android developer tools and performance at Android Dev Summit (Android Developers Blog, October 28, 2026)
• ✦Developing for Today: 50+ New Developer Sessions on Building Apps Today (Android Developers Blog, November 2, 2026)
• ✦Android Studio Rabbit 2 release notes