Realtime AI News
Google rewires Android development around agents, opening app functions to Gemini
At Android Dev Summit 2026 on October 28, Google updated Android Studio, the Android CLI, its profiling tools and Jetpack Compose, and introduced AppFunctions, Android MCP and A2UI. Developers can now bring any ACP-compatible coding agent into the IDE, and expose app features to system-level AI entry points through a unified interface.
Google used its Android Dev Summit 2026 on October 28 to push out a broad set of developer tools and platform updates spanning Android Studio, the Android CLI, profiling tooling and Jetpack Compose.
The through-line is making AI more than an autocomplete layer — letting agents take part in building, testing, debugging and tuning a project, while leaving model and tool choice to developers. Google is not standardising on one model; it is exposing platform knowledge, build and test capability, real devices and profiling to whichever agent a developer prefers.
The most direct change is Android Studio Rabbit 2's stable support for Bring Your Own Agent. Developers can connect the coding agent of their choice inside the IDE; Google says more than 40 ACP-compatible agents are supported, including Claude Agent, OpenAI Codex and Google Antigravity.
Connected agents inherit Android Studio's own capabilities — project structure, symbol resolution, Compose previews, emulators and Logcat. A new Change Review feature audits agent edits or Git changes, and a /review command lets an agent inspect a larger diff and keep fixing what it finds. Android Bench 2.0 scores model-and-agent combinations on real Android tasks such as building an app from a design or migrating UI, checking whether the feature actually works, not just whether it compiles, plus time and cost.
The Android CLI pushes the same capability into terminals and CI. android device remote requests physical devices from Google's device lab, Android Device Hub manages local, remote and emulated devices in one place, and android preview renders Compose previews straight in the terminal. In internal tests, Google says agents using that command finished UI tasks 77% faster with up to two-thirds fewer tokens.
Quality work is next. Android CLI integrates LeakCanary for heap-dump analysis, Android Performance Analyzer can be driven from the CLI to chase cold-start latency, jank and background wakeups, and android signals pulls Android vitals and Firebase Crashlytics production data into the terminal. Jetpack Compose 1.13 adds recomposition tracing. Google says stronger models running the leak-hunt-to-fix loop via CLI were nearly twice as fast with up to 25% fewer tokens.
For how apps meet AI, Google announced three interfaces. AppFunctions, in Beta, is Android's unified tool-calling layer for exposing app features to system-level or in-app agents. Android MCP connects app capabilities to different agents, with a Connected Apps Platform exploring links into system-level entry points such as Gemini. A2UI, in Alpha, lets agents render replies as interactive interfaces through Compose's native renderer, using the app's own components.
On-device work continues as well: ML Kit gains an agent API built on Gemini Nano 4, in Alpha, with function calling, conversation history and execution-flow control, while the emulator for XR, audio and display glasses reached stability.
Maturity varies — some pieces are stable, others Alpha, Beta or Canary — so gains will depend on the project. The real test is twofold: whether agents can reliably call tools, hold project context and verify their own edits, and whether apps start exposing functions to AI entry points instead of serving only people tapping a screen.
Why it matters
By keeping model choice open and exposing app functions to multiple agents, Google is upgrading day-to-day developer tooling now while laying track for a future where users ask an AI and an agent operates the app.
Nearby Updates
All10/29, 10:57
FuriosaAI raises $620M Series D to scale AI inference silicon
FuriosaAI closed a $620 million Series D round on October 28, with the South Korean chip developer saying the money will meet surging enterprise AI inference demand and speed up next-generation silicon. Its RNGD accelerators and NXT systems target the inference layer, where buyers are more open to alternatives to Nvidia.
10/29, 10:18
Diffuspace ships 350B-parameter Clover, a diffusion model built for agent work
Diffuspace released Clover-35B-A3B-Preview, a diffusion language model with 35 billion total parameters and about 3 billion active per token that writes code, builds web pages and chains tool calls. On the 49-task WindTunnel web agent benchmark its median run took 1.52 seconds, less than half the 3.21 seconds recorded by the Jev and Mercury 2.5 combination.
10/29, 11:40
Reuters: researchers say a Chinese AI model ran complex cyberattacks in controlled tests
Reuters reported on Oct. 29 that researchers observed a Chinese AI model carrying out complex cyberattacks inside controlled tests. The report describes an evaluation finding rather than a confirmed real-world intrusion, and the underlying details still depend on the full article.
10/29, 09:00
The Information: OpenAI's Astra becomes the top distillation target for Chinese AI labs
OpenAI's Astra model has become the primary distillation target for Chinese AI labs, according to The Information, which names it as the source those teams reach for first. The report identifies the target of the effort rather than naming the labs involved or quantifying its scale.