A2UI Agent SDK

Helps an AI agent generate A2UI: it negotiates catalogs with the renderer, teaches the LLM the catalog through a system prompt snippet, and parses the LLM's response into A2UI messages.

The package follows the agent SDK blueprint, limited to A2UI protocol v0.9 and the Express inference format.

Status

Implemented for one agent turn:

• A2uiGenerator.createProcessor picks the registered catalogs the renderer supports.
• A2uiRequestProcessor.promptSnippet teaches the model the Express syntax and the catalogs' components and functions.
• A2uiRequestProcessor.parseResponse compiles each Express block into v0.9 messages and checks them the way a renderer would.

Not supported yet:

• Protocol versions other than v0.9, and formats other than Express.
• Standalone function calls such as openUrl("...") on a line of their own, since v0.9 has no message for them.
• Catalog transformers and prompt examples.
• Decompiling messages into Express, and Parser.wrap.
• Function descriptions in the prompt, since a2ui_core's FunctionApi does not keep them.

Usage

See example/a2ui_agent_example.dart for one agent turn:

final generator = A2uiGenerator(
  catalogs: [CatalogConfig(catalog)],
  inferenceFormatFactory: const ExpressFormatFactory(),
);
final A2uiRequestProcessor processor = generator.createProcessor(
  rendererCapabilities,
);
final String llmOutput = await callLlm(processor.promptSnippet, userMessage);
final List<ResponsePart> parts = processor.parseResponse(llmOutput);

Tests

test/conformance runs the shared suites under conformance/agent/express. They are written against v1.0, so the harness lifts the v0.9 messages this package emits into the v1.0 shape before comparing. Cases for what is not implemented yet are skipped with the reason.

e2e_test runs the example turn against a real Gemini model. It needs an API key, so it lives in a separate package and runs in a separate CI pipeline.