Trace
The replayable record of a single agent run: prompts, context, tool calls, outputs. The black box recorder.
A trace is the full, replayable capture of one agent run: what it was asked, what context it saw, every tool it called, and everything it produced along the way.
Why it matters
When an agent does something wrong, a log line saying “error” tells you nothing. You need to see the run. What did it think it was doing? Which tool call went sideways? What context misled it? Without the trace, every incident is a guess and every fix is luck.
In practice
Each run emits a structured trace you can open, read step by step, and replay. Good traces feed directly into evals: a bad run becomes a test case. They are the raw material for both debugging and improvement.
Where it shows up in my work
The trace is where monitoring starts and assurance begins. It is the first artifact in the failure to assurance loop: failure, trace, eval, policy, remediation, proof.