AI 開發工具
Android Studio Quail 4 Stable Ships with 23 Agent Skills, Gemma 4 Enables Offline Cross-File Refactoring
Google is bringing automatic selection of Android-specific skills and on-device Gemma 4 inference to Android Studio’s stable channel. Code can remain on the developer’s machine, but local execution does not mean the agent has robust permission isolation or verifiable quality.
Google has released the stable version of Android Studio Quail 4 (2026.1.4), consolidating agent capabilities previously tested in stages into workflows suitable for production development. The IDE comes with 23 Agent Skills written by the Android team, covering tasks such as upgrading to Android Gradle Plugin 9, diagnosing performance with Profiler, migrating to Navigation 3, and building adaptive interfaces. After receiving a prompt, the agent matches it against skill metadata and loads instructions as needed. Teams can also add their own skills, while the same official skills can be installed through Android CLI for use with other command-line agents.
Another key change is that the local Gemma 4 model is no longer limited to preview. Developers can download the model through the model selector, while Android Studio handles verification, weight updates, and execution through its built-in lightweight inference engine. Gemma 4’s tool-calling capabilities support planning, modifying, and inspecting multiple files. In offline mode, source code does not need to be sent to the cloud, and usage is not constrained by API token quotas. Google says the smallest model can run on a system with 12GB of RAM, though it recommends at least 32GB. This means context capacity, speed, and quantization quality for large projects will still be constrained by workstation hardware.
Quail 4 also improves status indicators for parallel agents, adds clickable code symbols, and introduces a Summary of Changes that displays diffs in one place. Engineering teams should still treat modifications generated by local models as inputs requiring review: Google has not published cross-project refactoring success rates, test pass rates, or controlled comparisons with cloud models, and local execution does not demonstrate that tool operations are protected by strong sandboxing. Key issues to watch next are whether skill versions can keep pace with rapidly changing Android APIs and whether enterprises can pin model, skill, and inference-engine versions to establish reproducible CI validation workflows.