What is Agent Tracing?

Because an agent's final answer is the product of many hidden steps, a tool call here, a retry there, a piece of retrieved context somewhere else, debugging failures purely from the end output is nearly impossible. Tracing captures each step as a discrete span with timing, inputs, and outputs, typically nested to reflect parent-child relationships between an orchestrating agent and any sub-agents or tool invocations it spawned. Traces serve two audiences: engineers diagnosing why an agent produced a wrong or unsafe result, and compliance teams needing an audit trail of what data an agent touched and what actions it took. Mature tracing setups tag each span with structured fields, stage, tenant, session, result, and reason, so failures can be searched and aggregated rather than read one transcript at a time.