AG-UI brings A2UI Express into ADK and defines recovery boundaries for generation errors, cancellation and disconnects

AG-UI released ag-ui-a2ui-toolkit 0.0.6 on November 2 and ag_ui_adk 0.12.0 on November 3. The toolkit feeds structured A2UI generation errors into the next bounded model attempt and adds an optional Express inference format. ADK can opt into Express per run, incrementally compile it into validated activity updates, and clean up model and producer streams across cancellation, HTTP disconnects and backpressure.

Why it matters

The coordinated releases move generative UI reliability beyond producing valid JSON to how validation failures are fed back, how UI identity survives incremental compilation, and how producers unwind when a client disappears.

A closer look

Toolkit 0.0.6 records structured A2UIGenerationError data and returns it to the model on the next attempt within the existing attempt limit; transport, cancellation and other non-generation exceptions still propagate. It adds inference_format with direct_json and express options, retaining direct_json as the default. ADK 0.12.0 enables per-run A2UI Express for injected or owned render tools. A2UI agent SDK 0.10 incrementally compiles Express into validated activity updates with stable surface and activity identity. Cancellation awaits model-stream unwind. HTTP disconnects cancel and await the producer even under backpressure, while preserving iterator cleanup and HITL pauses; the SSE close path covers sse-starlette 2.1.0 and later.

Who this is for

Engineering teams building AG-UI, A2UI, ADK streaming renderers and human-in-the-loop workflows.

Original sources

• AG-UI official immutable November 2 and November 3 releases ↗github.com
• AG-UI Release 2026-11-02 ↗github.com