Eval and routing patterns (advanced)

Production agents fail when the wrong tool returns a fluent answer. Agent KBase does not ship a router, but the same discipline applies:

Golden set

Fixed inputs with expected outcomes. Fail the build when routing or retrieval regresses.

Tracing

Observe decisions after deploy (Langfuse, Langsmith, etc.). Demos are not traces.

Deterministic hot paths

Prefer code for tool choice; add model tie breaks only with eval proof and a latency budget.

Wizard as eval gate

The web wizard blocks export until required fields pass validateBrainInputStrict. Same idea: no ship without checks.