Android Studio now lets you connect other agents alongside the built-in one. The feature is called Bring Your Own Agent (BYOA). Google's announcement, posted September 24, 2026, walks through plugging Claude Agent, Codex or Antigravity into the IDE. For an Android team lead, the decision that matters is which agent to connect, what that agent can reach once it's inside your project, and how much risk is still yours to manage: Google describes granular permissions only in general terms and publishes no configurable policy.
What is Bring Your Own Agent, and why did Google build it this way?
BYOA runs on the Agent Client Protocol (ACP), and that choice explains why Google didn't just bake in one more assistant. ACP's own introduction frames the problem as an integration tax: without a shared protocol, every editor needs a custom integration for every agent, and every agent needs editor-specific code to run anywhere. ACP standardizes that handshake roughly the way the Language Server Protocol standardized editor-to-language-server communication years ago. Wire an editor to ACP once, and any ACP-compliant agent works with it. That's also why Android Studio can offer three unrelated agents, one each from Anthropic, OpenAI and Google, through a single settings panel instead of three separate plugins.
HighCircl's AI in the engineering workflow: tools, permissions and policy hub collects our coverage of how other tools have handled agent permission policies. BYOA's permission model, covered below, is the newest entry in that story.
Which agents can you plug in today?
Three agents connect without any setup beyond signing in. A fourth path exists for everything else.
Claude Agent (Anthropic)
Claude Agent is Anthropic's entry. Teams already running Anthropic's Claude Code may have met a related change: Claude Code recently added support for AGENTS.md, the shared agent-configuration file. Google's announcement doesn't say whether that file applies inside Android Studio.
Codex (OpenAI)
Codex is the second default. Google's announcement and Android Studio's preview features page both list it alongside Claude Agent and Antigravity as one of the three agents available without registering anything yourself.
Antigravity (Google)
Antigravity is Google's own agent, and its track record on real Android work is already public. On Google's own benchmark of AI agents against long-horizon Android tasks, the Antigravity-paired Gemini 3.7 Flash and Gemini 3.8 Flash scored 7.3% and 8.0% overall pass rates across the benchmark's 30 long-horizon tasks. Antigravity showing up as a BYOA default doesn't automatically inherit that result (different task, different pairing), but it's the closest public data point on how a Google-agent setup handles Android-specific engineering.
Other agents via the Settings registry
Anything outside the three defaults registers through File (Android Studio on macOS) > Settings > Tools > AI > Agents, an open slot for additional ACP-compliant agents.
What does a connected agent get access to inside Android Studio?
Google's announcement lists a wide surface once an agent is connected: the full project graph, build diagnostics, Jetpack Compose Previews, Android SDK tools, emulator control, file reads, writes and edits, shell command execution, test running, web search and the Android Knowledge Base.
How does the permission model work, and what isn't published yet?
Google's own language for the permission model stays general. The announcement says BYOA ships with granular permissions that let an agent work through routine tasks without asking, while pausing for approval when it hits something riskier. Neither the announcement nor the preview features page defines what counts as routine versus risky, and neither publishes a configurable matrix an admin can set and lock down.
That's a real gap next to what two adjacent tools already ship. GitHub's deny, ask, allow policy for Copilot agent operations lets an admin set a three-state rule per surface (shell commands, file edits, network domains) that a developer's own settings can't override. GitLab's default permission policy for Duo takes a different default (read-only tools allowed automatically; write and delete actions requiring approval), but it's still an explicit, documented, admin-adjustable policy. Android Studio's BYOA has neither shape yet: no published matrix, no described admin-policy layer. For now, whatever risk tolerance you want for a connected agent is a decision you make and monitor yourself.
How to set up Bring Your Own Agent in Android Studio
1. Update to the Rabbit 2 canary channel
BYOA ships in Android Studio's Rabbit 2 Canary 2 release, so update your canary channel before looking for it. It isn't in the stable channel, and the preview features page gives no timeline for when, or if, that changes.
2. Connect a default agent
Open the agent window and pick Claude Agent, Codex or Antigravity. Sign in or paste an API key for whichever one you choose. That's the entire setup for the three default options, with no separate registration step.
3. Register additional agents
For anything outside the three defaults, go to File (Android Studio on macOS) > Settings > Tools > AI > Agents and register an ACP-compliant agent there. Both sources describe this as the path for agents Google didn't ship as a default.
Should a team run this on a canary build with a real project?
Rabbit 2 Canary 2 is preview software, and BYOA sits inside it, so running BYOA means running unreleased Android Studio at the same time as an agent with shell access, emulator control and write access to your files. Neither risk is new by itself. Canary builds break in ways stable releases don't, and agents with broad access carry their own failure modes. Running both together on a real, shared codebase is where those risks compound, and there's no published admin-policy layer yet to cap the damage if something goes wrong.
A reasonable rollout looks less like flipping BYOA on for the team and more like one engineer, on a branch, on a project you can afford to lose a day rebuilding. Watch what the agent actually touches (shell commands, file writes, emulator actions) before deciding whether the routine-versus-risky split Google describes matches what you'd actually call routine.
What's available for enterprise and Gemini Enterprise plans?
Google's announcement says BYOA supports both consumer and enterprise AI plans. Gemini Enterprise users specifically get the built-in agent or Antigravity with the security and privacy of Google Cloud, in the announcement's words. What the announcement doesn't describe is a BYOA-specific admin policy: no separate console setting for which of the three agents a team can connect, no org-wide deny, ask, allow rule like GitHub's or GitLab's. For now, enterprise readiness means the surrounding Google Cloud controls apply.
FAQ
What canary channel is Bring Your Own Agent available in?
Android Studio Rabbit 2 Canary 2. It isn't available in the stable channel.
Which AI agents work with Android Studio's Bring Your Own Agent?
Three connect by default: Claude Agent from Anthropic, Codex from OpenAI, and Antigravity from Google. Any other ACP-compliant agent can be registered through File (Android Studio on macOS) > Settings > Tools > AI > Agents.
Does Bring Your Own Agent work with Gemini Enterprise plans?
Yes. Gemini Enterprise users can use the built-in agent or Antigravity under Google Cloud's existing security controls, according to Google's announcement. There's no published BYOA-specific admin policy layer on top of that.
Is the Agent Client Protocol specific to Android Studio?
No. ACP is an open protocol, built the way the Language Server Protocol was, as a shared interface any editor and any agent can implement. Android Studio is one adopter, not the protocol's origin.