Google used its Android Dev Summit on October 28, 2026 to hand the Android development environment over to outside artificial intelligence agents. The headline change is Android Studio Rabbit 2 with a stable Bring Your Own Agent feature, giving developers access to more than 40 compatible coding assistants without leaving the IDE. Anthropic’s Claude Agent, OpenAI’s Codex, and Google’s Antigravity can run side by side through a built-in marketplace. This is a meaningful change for Android Studio agents because they can inspect project structure, resolve symbols, operate the emulator, read Logcat, modify code, and verify results against a live application rather than simply offering suggestions in a chat panel.
The practical effect is that a developer can choose a model for a particular task, keep several agents active, and let them work against the same development context. One assistant might handle a migration while another investigates a failing test. A third could check whether a generated change introduced a regression. Because the agents share Android-specific tools, the competition is shifting away from a single assistant imposed on every project and toward a more measurable contest between models, prompts, and workflows.
Android Studio agents gain direct access to the development environment
The stable Bring Your Own Agent capability uses the open Agent Client Protocol, or ACP, as the common layer for connecting assistants to Android Studio. Google says more than 40 agents support the protocol. The key benefit is not simply that code can be pasted into the editor. Connected agents can query the project, locate symbols, navigate the codebase, launch the emulator and examine logs. That makes the IDE a shared execution environment rather than a text box wrapped around an assistant.
Several safeguards and companion tools were also introduced. Change Review analyzes pending modifications, prepares bug reports and explains changes step by step, which could make it easier for a developer to audit work produced by machine. Android Studio Canary includes a migration assistant for converting older XML layouts into Jetpack Compose. Google is also testing Lightbuild, a declarative build system based on a YAML schema that people and agents can both read and edit. The company also released version 2.0 of Android Bench, a leaderboard that evaluates models on multi-day development projects by functionality, completion time and cost.
These additions suggest that Google is preparing for a workflow in which agents make more than isolated edits. They may plan a change, run tests, inspect output and revise their work. As a result, review, reproducibility and cost control become as important as raw code-generation quality. An agent that finishes quickly but introduces subtle defects is not useful on a complex Android project. A slower system that can show why it changed a build file or why a device test failed may be more valuable.
The Android CLI brings the same capabilities to terminals and CI
Developers who run assistants in a terminal or automated pipeline receive a similar expansion through the Android CLI. Google released it as an official plugin for Claude Code, Codex and Antigravity. New commands allow an agent to render Jetpack Compose previews directly in the terminal and inspect a user interface change without starting an emulator. According to Google’s internal testing, agents completed interface tasks 77 percent faster and used as many as three times fewer tokens when they could use the preview capability.
That figure is a company-provided result, not an independently verified guarantee for every project. It still indicates why visual feedback matters in agentic development. A code generator can modify a layout correctly according to source code, but it cannot know whether labels overlap or controls fall outside the usable area. Rendering the preview gives it another signal, reducing the need for a full emulator cycle and potentially preserving more tokens for substantive reasoning.
The CLI also supports streaming physical test devices from Google labs, allowing an agent to investigate failures tied to particular hardware. Other commands retrieve crash information from Firebase Crashlytics, capture traces through the Android Performance Analyzer and examine optimization through the R8 compiler. LeakCanary is also integrated into both the CLI and Android Studio Profiler. Its heap-dump analysis now runs on the computer instead of the device, reducing test-time interference with the app. In a Google A/B test using Claude Opus 5.5, the CLI approach found and repaired memory leaks almost twice as quickly while using roughly 25 percent fewer tokens.
Again, that is Google’s own comparison rather than a universal benchmark. Its value is in showing where structured tooling can outperform a generic coding loop. An agent that receives a crash report, a performance trace or a heap analysis is working from evidence that is difficult for a language model to infer from source code alone. The command line also makes these capabilities available in CI, where a developer may want an agent to investigate failures, summarize diagnostics or propose a narrowly scoped repair.
Android Skills address the limits of model training
One persistent problem with coding agents is that newer interfaces may not be well represented in a model’s training data. Google is releasing more than 30 Android Skills that provide relevant documentation to agents. Their examples include checking Google Play policies before submission, handling permission security concerns, and adapting layouts for foldables and laptops branded as Googlebooks. The skills are intended to give agents targeted operating knowledge instead of expecting them to infer current Android conventions from general training material.
This is especially important for mobile development, where an apparently valid code change can still fail because a permission model, policy requirement or adaptive layout rule has changed. Google also updated Jetpack Compose 1.13 and the AppState model to support phones, large displays, foldables and Googlebook-style laptops with a more adaptive approach. An agent equipped with current instructions could help developers modify one codebase for several screen configurations, rather than treating each form factor as a separate application.
The emphasis on documentation and controlled tools shows that access to an IDE alone does not make an agent reliable. Models need current platform rules, valid interfaces and ways to test their assumptions. Google is building those layers alongside the agent integrations, which may make Android Studio more useful to small teams that cannot maintain extensive internal prompts, test harnesses and documentation systems for every project.
Agentic APIs move the concept into applications
Google’s announcements also reach beyond developer tooling. An alpha Agentic API powered by Gemini Nano 4 is being added to ML Kit for on-device processing. It can execute functions, retain chat history, manage the agent loop and apply safety filters. A beta AppFunctions API and support for the Model Context Protocol create a common tool-calling layer through which applications can expose capabilities to local or system-wide agents.
An alpha Agent-to-UI interface for Jetpack Compose adds another layer. Instead of returning a plain text response, an agent could present interactive controls built from the host application’s components. That distinction matters because an agent that can create a usable interface may require less effort from the user than one that describes an action or produces raw markup. The app retains control of its design system and component behavior while gaining a more capable way to express a response.
Google also previewed the Connected Apps Platform, which allows developers to connect applications through Android MCP or cloud-hosted MCP servers and manage the connections in one console. More than 50 integrations from companies including Adobe, Peloton and Experian are already available in the Gemini app, according to Google, with planned expansion into web search. Google did not provide a fixed general-availability schedule, so developers should treat the platform and its distribution plans as preview-stage details that may change.
Android XR moves closer to stable development
The summit also marked a broader shift in Android’s tooling. The Core Jetpack XR SDK libraries, Android XR emulators and the XR Engine Hub for Unity, Godot and Unreal have reached stable releases. Geospatial capabilities in ARCore were added to the development workflow, allowing digital content to be anchored to real-world locations. Emulators can help teams test XR applications without maintaining dedicated hardware, while performance traces remain important because an experience that works in a simple demo may not perform well with realistic content.
Google’s stable XR tools may eventually make Android Studio agents more useful in spatial interfaces. An agent that can edit an XR scene, read diagnostics and inspect a preview would need the same project-aware access that is becoming standard for conventional Android apps. This announcement is not yet proof that autonomous XR development is dependable, but it broadens the environments in which the same agent-centered toolchain can be tested.
XREAL AURA brings Android XR to wired glasses
Hardware partner XREAL introduced the AURA, which XREAL executive Shahram Izadi described as the first wired smart glasses running Android XR. The glasses weigh less than 95 grams and use transparent dimmable lenses with a 70-degree field of view. Their computing unit and battery reside in an external puck that also functions as a trackpad and runs on Qualcomm’s Snapdragon Reality Elite chip.
The AURA can place as many as five Android applications in the user’s field of view. Gemini can interpret the physical room together with the virtual screens, allowing it to organize windows, suggest actions during a game or highlight objects in the environment. The device can also mirror a laptop through USB-C. It is available to order in the United States, Canada and South Korea for 1,099 dollars, with Google and XREAL saying shipping will begin in mid-November.
For developers, the glasses illustrate why Android Studio agents need more than code completion. Spatially aware applications must respond to context, preserve battery life and behave well across a wide field of view. Agents that can operate project files and performance tools may help lower some of that burden, but visual verification remains essential because an agent cannot infer every spatial interaction from code.
The strategic shift is a shared, tool-aware workspace
Google’s central move at the summit was to stop presenting its own coding assistant as the only route into Android development. With more than 40 agents able to share the emulator, logs and project context, the IDE is becoming an operating layer for several models at once. The command line extends that approach into terminals and CI, while Skills, Change Review and specialized diagnostic tools address reliability. Whether that model choice actually reshapes how Android software gets built, and how far the Android Studio agents workflow travels beyond Google’s own tools, will decide how much this release changes the platform’s centre of gravity.
That does not remove the need for human judgment. Generated code still requires review, security policies still matter, and company-reported speed improvements may differ across repositories. It does, however, change the central question for Android Studio agents. The important competition is no longer simply which assistant a developer prefers, but which agent can complete a real task on a real Android project, prove what it changed, and work safely inside the tools the team already uses.