tensorfolk

Google's A2UI agent SDKs add swappable output formats, with compact A2UI Express as the new default

Google updated the A2UI generative-UI SDKs so developers pick how a model writes interface code with a single generator argument. The new default, A2UI Express, is a Python-like syntax compiled into standard A2UI JSON, which Google says cuts tokens and latency by up to 3x. Python 0.9.0 also adds macros and a generator that turns a component catalog into agent skills.

Why it matters

Agent builders emitting UI can cut latency and token spend without changing renderers, since Express compiles back to standard A2UI messages.

Source: Read the full story on Google Developers Blog

a2uigenerative-uiagent-sdkgoogle

More in AI Tools & Open Source

AI Tools & Open Source / Frameworks & SDKs

Modular 26.7 publishes all of MAX's source under a new source-available license

Modular released the remaining MAX code, including its graph compiler, runtime and hardware backends, joining kernels and serving code already under Apache 2.0. The new MAX License permits free commercial use on any hardware but bans distributing modified forks and using the code to train AI that reimplements it.

Why it matters: The whole MAX inference stack is now inspectable, but it is not OSI open source, so forks are restricted.

Read on Modular
AI Tools & Open Source / Coding Assistants

GitHub makes Copilot Fusion multi-model routing generally available at Universe 2026

At GitHub Universe, Copilot Fusion left research preview and became available on every Copilot plan. Picking Fusion in the model picker hands each prompt to a router that decides which models to call and how to combine them, with Standard, High and Max effort levels. GitHub says Fusion Max lands near top single models on a coding-agent index at roughly half the cost.

Why it matters: Teams can stop hand-picking models per task; admins still control which models Fusion may route to, and billing stays per underlying model.

How to try it

Copilot Fusion is a routing system you select instead of a single model; it works in VS Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs on all Copilot plans, billed at each routed model's rate.

  1. Update VS Code (1.144+), the GitHub Copilot app, Copilot CLI or your JetBrains Copilot plugin.
  2. Open the Copilot model picker and select Fusion.
  3. Choose an effort level: Standard, High or Max.
  4. Org admins: set which models Fusion may use via enterprise/organization model policies, and optionally make Fusion the default via enterprise managed settings.
Open the docs
Big newsRead on The GitHub Blog
AI Tools & Open Source / Coding Assistants

Android Dev Summit: Android CLI ships as official Claude Code and Codex plugin; bring-your-own-agent goes stable

Google's Android tooling update makes the Android CLI an installable plugin for Claude Code, Codex and Antigravity, giving agents commands for Compose previews without an emulator, remote real devices, profiling and Play-style bundle analysis. Android Studio Rabbit 2 also makes its Agent Client Protocol support stable, so 40+ third-party coding agents can run inside the IDE.

Why it matters: Android teams can use their preferred coding agent with first-party device, preview and profiling tools; Google reports 77% faster UI tasks with fewer tokens.

Read on Android Developers Blog