Tag: observability

A trace records how an agent reached an answer, including model calls, tool use, handoffs, retries, and verification.

A practical loop for turning recurring agent failures into evidence-backed eval definitions.