traces

Definition

Step-by-step records of what a system did: the agent's reasoning, its tool calls, and the environment's responses, kept in order. Traces are how agent failures get diagnosed at all — a final wrong answer says nothing about where the run went off — and they double as training data, since filtering for successful traces yields supervision without human annotation. The value, as in distributed tracing, is preserved causal ordering rather than isolated log lines.

Explainers using this term

• Paper Walkthrough: AutoSaddler — Growing a Harness That Doesn't Break, from Agent Failure LogsAutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces