A reasoning trace is the recorded sequence of intermediate thoughts, tool calls, and decisions a model produces while solving a task, distinct from its final output. It serves as both a working scratchpad that improves accuracy and an audit artefact for inspecting how an agent reached a conclusion. Traces underpin debugging, evaluation, and self-correction in agentic systems.
Content
- Traces capture chain-of-thought text, tool invocations, observations, and branch points, enabling replay and inspection of agent behaviour. They are central to evaluation and safety, since a misleading or unfaithful trace can mask the true cause of an output.