Beyond the News
Sign in
Artificial IntelligenceWeb DevelopmentAI CodingDeveloper Tools Active dossier

Android Studio Quail 4 makes curated agent skills and offline Gemma 4 production-stable

Google’s final stable Quail release bundles 23 Android-specific agent skills directly into the IDE and adds native Gemma 4 inference for fully offline multi-file coding. The same skills can also be installed into external CLI agents through Android CLI.

Reviewed and published by Lars Jansen Updated 6 min read
Comment
First published

Key details

  1. Android Studio Quail 4 became stable on September 1, 2026 and is the final stable release in the Quail series.
  2. The IDE ships with 23 curated Android agent skills following the open Agent Skills specification.
  3. Android Studio automatically searches installed skill metadata and invokes relevant skills based on the developer's prompt.
  4. Android CLI can install all of Google's Android skills for compatible external CLI agents using `android skills add --all`.
  5. Android Studio natively integrates Gemma 4 with downloadable model management and a bundled local inference engine.
  6. Google says Gemma 4 can run multi-file tool-calling agent tasks completely offline so source code stays on the local machine.
  7. Google lists 12GB RAM as sufficient for the smallest local configurations and recommends 32GB+ for the best experience.
  8. Quail 4 also adds background-agent status indicators and a unified Summary of Changes review surface.

What builders should take away

  1. If an agent repeatedly produces stale Android patterns, use maintained skills as a grounding layer before reaching for larger prompts or a more expensive model.
  2. Treat reusable agent instructions like code: version team-specific skills, review them when platform conventions change and keep them close to the workflows they govern.
  3. For sensitive repositories, test the Gemma 4 local path separately from cloud Gemini, third-party provider and MCP configurations; only the local model path carries the offline/source-stays-local guarantee described by Google.
  4. Size developer hardware for the local model you actually intend to run. A nominal 12GB floor does not imply every useful configuration will perform well there.
  5. Because Google's Android skills are portable, teams using multiple coding agents can standardize platform guidance once rather than duplicating tool-specific prompts.
  6. Use the Summary of Changes and background status UX as review boundaries; parallelism increases throughput but also increases the amount of agent output a developer must validate.

What changed

Android Studio Quail 4 became the final stable release in the Quail series on September 1, 2026. Google now ships 23 curated Android agent skills directly in the IDE, following the open Agent Skills specification. Android Studio searches installed skill metadata when a developer prompts the agent and automatically invokes relevant guidance for workflows such as Android Gradle Plugin upgrades, profiling, Navigation3 migration and adaptive UI work. Google has also made native Gemma 4 integration part of the stable IDE: developers can download and manage model weights from Android Studio, use a bundled local inference engine and run multi-file agentic coding tasks offline. The Android skills are not locked to the IDE; Android CLI can install the full set for compatible external command-line agents with `android skills add --all`.

Why it matters

The release turns two ideas that have mostly lived as add-ons or previews — reusable agent instructions and local coding models — into first-class parts of a mainstream development environment. For Android teams, this creates a maintained path for grounding coding agents in fast-changing platform conventions instead of relying entirely on a model's training data or long repository prompts. It also gives teams with source-code privacy or connectivity constraints a supported offline agent path. The portability of the skills is equally important: Google's Android-specific instructions can travel outside Android Studio, making the open skill format a practical interoperability layer rather than an IDE-only feature.

The IDE now ships with 23 platform-specific agent skills

Google bundles 23 curated skills designed around Android workflows. Android Studio matches prompts against skill metadata and invokes relevant instructions automatically, reducing the need for developers to manually attach platform guidance to every task. The initial set includes skills for AGP 9 upgrades, Android Profiler work, Navigation3 and adaptive layouts.

The skills use an open format and can leave Android Studio

The bundled instructions follow the open Agent Skills specification. Developers can also create custom skills for team-specific workflows, and Android CLI can install Google's full Android skill set for compatible CLI agents. That makes the material development broader than one Gemini integration: the domain knowledge is packaged as a portable agent extension.

Gemma 4 runs inside the stable IDE without cloud inference

Android Studio can download, verify and update Gemma 4 model weights and includes a lightweight local inference engine. Google says Gemma 4's tool-calling support allows complex multi-file refactoring plans to run completely offline, so source code stays on the local machine and cloud token quotas do not apply to that local path.

Local AI still has a real hardware floor

Google says the smallest supported Gemma configurations can run with 12GB of RAM, while machines with 32GB or more are recommended for the best experience. Local execution therefore trades cloud dependency for local memory and compute requirements rather than eliminating resource costs.

Parallel-agent work now has more explicit review UX

Quail 4 adds background status indicators for parallel chats and consolidates task output into a Summary of Changes view. The IDE can show when a background agent is running, waiting for input or ready for review, helping multi-agent work behave more like an observable development workflow than a collection of independent chat windows.

The stable release collects capabilities previewed across the Quail cycle

Some individual features appeared earlier in Canary or prior Quail releases, including agent skills, local Gemma support, parallel chats and an MCP marketplace. The September 1 event is the final stable boundary: Google now presents the combined agent stack as production-ready in the stable channel.

What to watch next

  • How often Google updates the bundled Android skills as APIs and recommended migrations change.
  • Whether the Android skill set is adopted broadly by non-Android-Studio coding agents through the open specification.
  • Independent comparisons of local Gemma 4 coding quality and latency against Android Studio's cloud-backed models.
  • Whether Google expands the local-model path to additional model families or hardware backends.
  • How the Android Studio MCP marketplace, portable skills and local models converge into a more composable agent ecosystem.

Still unclear

  • Many Quail capabilities were previewed earlier in the release cycle; the new material boundary is their inclusion in the final stable release rather than every feature being invented on September 1.
  • Google's claim that source code stays local applies to the offline Gemma path, not automatically to cloud models, third-party model providers or external MCP services configured in the IDE.
  • Actual local-model speed and useful context capacity depend on selected model size and developer hardware.
  • The practical portability of Android skills depends on how completely each external agent implements the Agent Skills specification.

Sources

Direct reading behind this dossier.

1 sources

Discussion

Discussion is reader-contributed. Comments are not part of the BTN dossier or its editorial evidence.

0 visible comments

Join the discussion

Keep comments useful and relevant. Reader contributions may be moderated and are not BTN editorial evidence.

Sign in to comment
Continue reading
Artificial Intelligence

Qwen3.8-27B brings stronger agentic coding into a locally deployable 27B model

Qwen3.8-27B is now available as open weights on Hugging Face and ModelScope. For builders, the important change is not another benchmark bump: a comparatively compact 27B model now combines native vision, long context, controllable reasoning and OpenAI-compatible serving paths for local or self-hosted coding and agent workloads.

Artificial Intelligence

OpenAI turns the Codex harness into a managed Agents API

The important shift is that agent orchestration itself becomes a managed API surface: context compaction, tool discovery, programmatic tool calls and subagent coordination can now come from OpenAI’s maintained Codex harness rather than an application team rebuilding those layers.