Playbooks · Coding tasks
Bring your own agent into Android Studio, and switch when its quota runs out
Written November 1, 2026
For an Android developer already paying for Claude or Codex: run that agent inside Android Studio with emulator and Logcat access, know what sign-in, quota, and data terms actually apply per agent, and switch mid-task without leaving the IDE.
What changed
At Android Dev Summit on 28 October 2026, Google announced that Bring Your Own Agent is now stable in Android Studio Rabbit 2. The Android Developers Blog post states that Android Studio now supports over 40 ACP-compatible coding agents directly inside Android Studio through Bring Your Own Agent and the open Agent Client Protocol (ACP), naming Claude Agent, Codex, Junie, and Google Antigravity. Google's release-notes page lists Android Studio Rabbit 2 in the Stable channel. The blog post says to open Settings, then Tools, then AI, then Agents, to browse and install agents from the marketplace, and that you can install more than one and switch between them as you work. A connected agent gets access to the understanding the IDE has of your project: project structure and build, symbol resolution, Compose Preview rendering, the Emulator, Logcat, and more. Bring Your Own Agent also brings Google Antigravity into Android Studio: logging in gives access to the latest Gemini models as they launch, with new Antigravity capabilities as they roll out. Rabbit 2 also adds Change Review, an agentic code review that starts from pending agent changes or the Git changes window, or from the /review command in the agent tool window, and Android Bench 2.0, a leaderboard that grades models and agent harnesses on long-running, real-world Android tasks by whether the solution actually works, and measures cost and time.
Two cautions belong in the same paragraph as the news. First, this feature was a Canary preview before it was stable, and setup text written for the preview described updating to the latest Canary release. Download from the Stable channel and confirm the channel before installing: if the setup instructions you are following mention Canary, stop and get the stable download instead. Second, the agents are not identical. Google's post names four agents and says any ACP-compliant agent works; it does not say all 40-plus agents behave the same way. The steps below cover only the agents Google names and the official plugins, and per-agent behavior, quota, and data terms must come from that agent's own provider.
Before you start
Download Android Studio Rabbit 2 from the Stable channel and confirm the channel in the download you have. Decide which agent you will connect first: the one you already pay for is the sensible first choice, because you already hold its subscription or API key. Before installing it, open that provider's own plan and terms pages and write down two facts: what quota or limit applies to the plan you hold, and what data terms apply to code and prompts you send through it. This page states no quota or data term for any agent, because those terms belong to each provider and change with the plan. If you cannot find them, do not assume the IDE changes them: connecting an agent inside Android Studio does not replace its provider's billing or terms.
Steps
- Install the agent from inside the IDE. In Android Studio, open Settings, then Tools, then AI, then Agents, and browse the marketplace. Install the agent you chose. You can install more than one: install a second agent now if you hold a second subscription, so switching later is a selection, not a new setup in the middle of a task.
- Sign in with that agent's own account. Authenticate with your personal or enterprise subscription for the agent, as the release notes describe for Bring Your Own Agent. Use the account that holds the plan you wrote down in "Before you start". If sign-in asks for a scope or permission you did not expect, stop and read the provider's terms before accepting: the IDE gives the agent project, build, emulator, and Logcat access, and the provider's terms decide what leaves your machine.
- Run one small task and watch what the agent can reach. Ask the agent to do one small, easily checked job in a copy of a project, such as explaining one file's structure or running one build, and read what it reports back. Confirm it can see the project structure and build information the blog post describes, and confirm the result yourself rather than taking the agent's word for it. This step tests the connection, not the agent's quality.
- Switch agents when the first one hits its limit. When the agent you are using reaches the quota or limit from its own plan, switch to another installed agent from the agent list and continue the same task. Tell the new agent, in plain words, what the previous agent did and what remains, because the new agent does not inherit the previous agent's conversation. Check the result again after the switch: a task finished by two agents needs the same reading of the diff and the same test run as a task finished by one.
- Use the Android CLI plugins when you do not need the IDE open. Android CLI is also available as an official agent plugin for Claude Code and Codex, and as an official Android plugin for Google Antigravity. For Claude Code, from a Claude Code session run
/plugin marketplace add android/android-cli-pluginand then/plugin install android-cli@android-cli-plugin; from the terminal, runclaude plugin marketplace add android/android-cli-pluginandclaude plugin install android-cli@android-cli-plugin. For Codex, from the terminal runcodex plugin marketplace add android/android-cli-pluginandcodex plugin add android-cli@android-cli-plugin. The plugin gives the agent tools to create Android projects, manage emulators and virtual devices, inspect live UI layouts, run Journeys, browse and install official Android skills, and search the official Android Knowledge Base. For Antigravity, install the Android plugin from Customizations, then Build with Google, in the Antigravity App. The same official Android skills are available from the command line: the blog post givesandroid skills listandandroid skills add []to install skills for any agent, andandroid docs searchto access the Knowledge Base. Google publishes over 30 skills across more than 20 areas, aimed at fast-moving areas where models often fall short, such as Compose performance optimizations, new and advanced Camera APIs, and security audits. - Review agent changes before they land. Use Change Review on pending agent changes, from the Git changes window or with
/reviewin the agent tool window, and read its report of potential issues before accepting the changes. Ask the agent to fix the issues it finds, then run the build and tests yourself. Change Review is a reading aid, not an approval: a clean report does not prove the change is right.
What can go wrong
- Installing the Canary build by mistake. Preview setup text described the Canary release. Rabbit 2 is stable. Confirm the Stable channel before installing, and check again if anything in the setup path mentions Canary.
- Treating all 40-plus agents as the same. Google names four agents and says any ACP-compliant agent works. Steps, quota, and data terms differ per agent. Only the named agents and the official plugins are covered here; anything else needs its own provider documentation.
- Assuming the IDE replaces the provider's terms. Quota, billing, and data handling come from the agent's provider, not from Android Studio. Switching agents inside the IDE switches which provider's terms apply. Read them per agent, before the task that depends on them.
- Losing the thread across a switch. A new agent starts without the old agent's conversation. State what was done, what was checked, and what remains, in the first message to the new agent, or the second half of the task will repeat or contradict the first.
- Trusting a preview speed figure as your result. In Google's internal testing, agents that used the
android previewcommand completed UI tasks 77% faster while using up to 3x fewer tokens. That is Google's internal test, on its own task set, not a promise for your project. Measure your own tasks if speed is the reason you switch.
Limits
- This page covers the IDE host and the official CLI plugins Google names. It does not cover the other announcements from the same summit, such as the agentic Android APIs, and it does not rate agents against each other: Android Bench 2.0 is the leaderboard for that comparison, and its results are outside this page.
- No per-agent quota, price, or data term is stated here. Each agent's provider page is the source for those facts, and a provider page that does not state them leaves them unstated here too.
Related playbooks
- Visual Studio 2026 runs on the Copilot SDK
- Let your agent build inside Roblox Studio
- Amp Spaces: one agent runs the rest for you
- Qwen Code 0.28 runs one shared agent loop
Sources
- https://techy101.com/2026/10/29/whats-new-in-android-developer-tools-and-performance-at-android-dev-summit-android-developers-blog/
- https://developer.android.com/studio/releases
Back to: Coding playbooks hub · Also: Onboard onto an unfamiliar codebase with an agent · Make sure your Claude Code skill actually loads when it should: test it in each agent before you rely on it