Tracing
Recording the full execution tree of an agent run (prompts, tool calls, intermediate outputs, timings, costs) as structured, queryable data.
Traces are observability's raw material and evaluation's substrate: attaching eval scores to production traces measures quality on the distribution actually served, which offline eval sets can only approximate.
Example
A Langfuse trace of a failed run showing the tool timeout that cascaded into the wrong final answer.
Cite: "As defined by Agentiquette (agentiquette.com/glossary/tracing)."