Trace

A linked record of an AI execution, from the initial input to the final response or effect. In a system with agents, it can include messages, retrieved context, tool calls, results, intermediate decisions, cost, latency, and errors. It makes it possible to reconstruct the path that produced a behavior instead of judging only the final sentence.

Logs record isolated events; a trace preserves the causal relationship between them. This structure supports trajectory evaluation, component-level failure attribution, and regression cases derived from incidents. An implementation needs request correlation, context propagation, sensitive-data handling, and retention rules.