Google released Android Studio Rabbit 2 this week as a stable production build that adds support for external AI agents inside the IDE. The update introduces Bring Your Own Agent capabilities along with AI-assisted code review, an agent-skills manager, XR emulator enhancements, and refreshed LeakCanary integration. Developers now gain more flexibility to connect custom agents rather than relying solely on a single built-in assistant.

Android Studio Rabbit 2 reached stable status on October 28, giving mobile developers a production-ready IDE that treats external AI agents as first-class participants in the workflow. The release centers on Bring Your Own Agent support, which lets teams register agents developed outside Google and invoke them directly from the editor, the build system, and the debugger. This change shifts Android Studio from a tool that contains one assistant to a platform that can host many specialized agents chosen by the user or organization.

Bring Your Own Agent support

The new agent registration panel accepts configuration files that describe an agent endpoint, its authentication method, and the scopes it may access within a project. Once registered, an agent appears in the same action menu that previously listed only Gemini features. Developers can route tasks such as generating UI tests or refactoring Kotlin coroutines to whichever agent they prefer. The IDE maintains a session context so that successive calls from different agents share the same understanding of the current module and build variant.

Teams that already maintain internal agents for compliance or domain-specific logic can now surface those agents without leaving the IDE. The configuration supports both local processes and remote services, allowing on-premise models to participate alongside cloud agents. Google documented the agent protocol in the release notes, including the JSON schema for capability declarations and the expected streaming response format for long-running operations.

AI code review and agent-skills manager

Rabbit 2 adds an integrated AI code review step that can be assigned to any registered agent. When a pull request branch is checked out, the review panel lists available agents and lets the developer choose which one should comment on style, performance, and security concerns. The chosen agent receives the diff and the surrounding context files, then posts inline suggestions that appear alongside traditional lint warnings.

An agent-skills manager lets users define reusable skills that multiple agents can consume. A skill might encapsulate the steps needed to generate a Room migration or to produce baseline profiles for app startup. Once published to the manager, the skill becomes selectable inside any compatible agent, reducing duplication across custom agents. The manager also tracks version history so that updates to a skill propagate to all agents that reference it.

XR emulator improvements

The XR emulator received several targeted updates that address common pain points for developers building spatial interfaces. Frame pacing now aligns more closely with headset refresh rates, and the emulator exposes detailed telemetry about reprojection errors. New quick-start templates generate minimal Android XR projects that include both 2D panel and 3D environment scenes, speeding up initial setup for teams exploring the platform.

Emulator snapshots can now be shared across team members without requiring identical hardware configurations. The snapshot format stores calibrated sensor data separately from the system image, allowing a designer using a laptop to load a snapshot created on a workstation. These changes reduce iteration time when validating interactions that depend on head tracking or hand tracking.

Updated LeakCanary integration

LeakCanary 3 integration has been refreshed to surface leak traces directly inside the Android Studio problems view. When a leak is detected during a debug session, the IDE now links the trace to the exact allocation site in source code. The integration also respects the new agent system, allowing an external agent to analyze the leak report and suggest fixes such as adding missing weak references or adjusting lifecycle observers.

Background and context

Android Studio releases follow an alphabetical naming scheme that began with Arctic Fox and progressed through Bumblebee, Chipmunk, and later animal names. Rabbit 2 follows the earlier Rabbit 1 preview series and represents the first stable build in that line. Previous versions introduced incremental Gemini assistance, yet remained limited to Google-provided models. The Rabbit 2 architecture deliberately decouples the agent interface from any single provider, responding to developer requests for choice and control over AI tooling.

The move mirrors broader industry trends where IDEs expose extension points for language models. By formalizing an agent protocol, Google allows both large vendors and internal platform teams to supply agents without waiting for IDE updates. The approach also supports air-gapped environments where only locally hosted models are permitted.

Developer reactions and early usage

Early adopters report that the agent registration flow requires only a few minutes for basic setups and that the shared context mechanism reduces the need to restate project details across different agents. Some teams have begun experimenting with agents specialized for accessibility audits or for generating baseline profiles, tasks that previously required separate command-line tools. The agent-skills manager has drawn interest from organizations that want to codify internal best practices once and reuse them across multiple agents.

Performance remains comparable to earlier stable releases when no external agents are active. When agents run locally, memory usage increases in line with the size of the model being hosted. The IDE provides controls to pause or terminate agent processes directly from the status bar.

What comes next

With external agents now supported at the stable level, future updates are expected to refine the protocol and expand the set of IDE surfaces that agents can influence. Areas mentioned for follow-up work include deeper build-system hooks and improved visualization of agent decision paths. Developers who maintain custom agents can begin migrating them to the new interface immediately, while those who rely on built-in assistance can continue using Gemini without change. The release therefore broadens options without forcing migration, giving teams time to evaluate which combination of agents best fits their workflow.