Agent Observability

An agent run is a nondeterministic multi-step program. When a user reports 'it did something weird', 'we reran it and it worked' is not debugging — it's surrender.

Capture the full trace: every model call (prompt hash, response, token counts, latency, model version) and every tool call (args, result, duration), structured as a span tree under one run ID. The trace — not the final answer — is the unit of debugging. Frameworks differ; the discipline doesn't.