Kamus AI
Execution trace
An ordered record of every agent step: messages, tool calls, inputs, results, errors and cost.
An execution trace is evidence of what the agent actually did, as opposed to the summary it writes at the end. When an agent claims it is done, match the claim against its trace: which commands ran and what they printed. Traces are also used to count cost and find wasted turns. Store them without secrets.
Also called: Execution trace, trace, agent transcript