An agent trace is the complete record of one run of an AI agent: the task it was given, every model call and tool call in order with arguments and results, the reasoning where it was captured, any human interventions and the outcome. It records what happened between the messages, which a chat transcript does not.
Why it matters when licensing company data
Tracing tools store a trace as a tree of spans: one span for the agent run, child spans for each model call and each tool call, with timestamps and token counts. The OpenTelemetry GenAI conventions name these invoke_agent, model and execute_tool spans, and keep prompt and tool content in attributes that are recorded only when content capture is switched on.
For anyone who reuses traces, completeness matters more than volume. A trace whose tool results were truncated, whose middle steps were sampled away or whose outcome was never recorded cannot show how the task was actually done or where the agent went wrong. Traces also carry whatever the tools returned, so secrets, personal data and customer content have to be removed before they leave the team.
See licensing agent traces from production, how to export agent traces and the agent trace redaction checklist.
Example
A finance team runs an agent that reconciles supplier invoices against purchase orders. Each night's run leaves one trace per invoice: the task, eleven model calls, the ERP lookups and their results, a note where a person approved an exception, and the final posted entry. When the team reviews its traces, it keeps only runs whose posting it can confirm in the ledger, removes vendor bank details from tool results and records the model and agent framework used for every run.
Source
Hugging Face Hub docs: agent traces. Definition reviewed November 9, 2026 by Origin Data Partners.