agent-platform-implementation

This skill should be used when a multi-agent build on Gemini Enterprise Agent Platform runs without erroring but is quietly wrong: "the agent invented data the tool returned correctly", "it works with curl but not in Gemini Enterprise", "my sub-agent returned 200 and the caller still made something up", "a tool argument is fabricated", "the orchestrator dropped half the list", "/run_sse returns 404", "how do I read a sub-agent's output", "how do I stop the model inventing tool arguments", "my eval score did not change", "eval grade scored the wrong run", "the metric scores 0 on a working agent", "my multi-turn eval case tests something that never happened", "the A2UI card renders but the button does nothing", "which A2UI components does Gemini Enterprise actually render", "one component of my surface is missing", "the model keeps re-reading its own surface", "how do I build an A2UI tree", or "a parallel stage produced no output and no error". Covers what actually crosses an agent-to-agent boundary, guards that a model cannot game, building an A2UI surface that survives the round trip, and the verification traps that report a score unrelated to your agent. Use agent-platform-architecture to choose a host or transport — including whether A2UI can render there at all — agent-platform-a2a for calls failing with a status code, and google-agents-cli-eval for eval methodology, dataset schema and metric choice.

Version

0.2.0