Google: AI Agents Using Android CLI Complete UI Tasks 77% Faster

At Android Dev Summit 2026, Google showed how to bring AI agents (programs built on language models that read a project, edit code and check the result on their own) into Android Studio, the Android CLI command line, a lab of real devices and the apps themselves. The 77% figure in the headline comes from Google's internal tests. On UI tasks, agents using the android preview command finish 77% faster and use up to three times fewer tokens.

A token is a chunk of text that a model uses to measure the amount of work and the cost of a request. The android preview command renders a preview of a Compose UI without an emulator and without an APK build.

Android Studio accepts third-party agents

Android Studio Rabbit 2 has reached the stable channel with a Bring Your Own Agent mode, which lets you connect your own agent. Any of more than forty agents that support the open ACP protocol will work. They include Claude Agent, Codex and Antigravity.

The agent gets what a human engineer normally uses from the development environment:

• project structure; • build data; • symbol resolution, meaning the links between names in code and their definitions; • Compose Preview rendering, meaning previews of Jetpack Compose UIs; • the emulator; • Logcat, the Android system log.

Change Review also made it into the stable release. It is an agent-driven review of edits: it finds problems in uncommitted changes and can fix them on its own.

Android CLI: a real phone and previews without a build

Android CLI works in any terminal and in CI, meaning automated builds on a server. It has several commands:

• android device remote provides a real phone from Google's device lab and connects it as a local device through adb; • android preview renders a preview of Compose UIs without an emulator or an APK build; this is the command behind the 77% figure; • the remaining commands cover profiling, crash reports from production, build analysis using Google Play's logic, and search of a knowledge base.

Google released official plugins for Claude Code and Codex for the first time. It updated the plugin for Antigravity. Alongside them, more than thirty skills (sets of instructions for an agent) shipped across two dozen topics, from permissions security to Google Play policies.

One separate example is the android profiler leakcanary command. It analyzes a memory dump on the computer rather than on the device, so the app does not slow down during analysis. According to Google's tests, the strongest models complete this scenario almost twice as fast and use up to 25% fewer tokens than without Android CLI. The figure is given for Claude Opus 5.5. The subcommand itself has only been promised so far.

Android Bench 2.0

Android Bench 2.0 is a leaderboard, meaning a table of results, for models and agent environments on long, real-world Android tasks. The tasks are sized for days of an engineer's work. Solutions are also scored on cost and time spent.

Agents inside apps

The second part of the announcements is about how agents work with apps on the phone.

• AppFunctions, a unified function-calling layer in Android, has moved to beta. A developer describes an app's capabilities, and a system or custom agent calls them directly without opening any screens. This is exposed externally through Android MCP.

• A2UI for Jetpack Compose is out in alpha. The agent responds with a live interface assembled from the app's own components rather than with a long block of text. • The agentic API in ML Kit on Gemini Nano 4, also in alpha, calls functions on its own, stores conversation history and sets guardrails.

• The Connected Apps Platform is open in preview. Through it, apps reach Gemini and Search. Integration goes through Android MCP or a remote MCP server built on the same AppFunctions APIs, and registration and testing take place in the Google Connected Console.

Search and Gemini now have more than fifty connected apps. Partners include Adobe, Calm, Peloton, Experian and Localiza.

What teams should keep in mind

The 77% and 25% figures come from inside Google. Some of the tools are still in alpha or preview, and the leakcanary subcommand has not been released yet. Before adopting any of this, run your own tasks on your own project.