How a trace is built
Every run writes one trace. Each step in the plan becomes a span with its inputs, its outputs, the tools it called and how long it took. Model calls and tool calls sit in the same tree, so you can see which decision led to which action.
Traces are written before a step's result is used. If a run stops halfway, the trace still shows every step that happened.
Replay a run
Open any finished run and step through it with the exact inputs each call received. Replays run against a snapshot of the tools, so nothing is sent or written a second time.
You can change one input and replay from that step to see what the agent would have done instead.
Compare two runs
Put a failing run next to a passing one. Trace aligns the two plans and marks the first step where the inputs, the outputs or the chosen tool differ.
Most failures show up in the plan before they show up in the result.
Approvals in the trace
When a step waits for a person, the wait is part of the trace: who was asked, when, what they saw and what they decided. A rejected step keeps its full context.
Retention and export
Traces are kept for 30 days on Starter and 400 days on Team and Enterprise. Export any trace as JSON, or stream them to your own store as they are written.
Access and audit
Traces can contain customer data, so access follows your workspace roles. Every time a trace is opened, that is written to the audit log.
Fields you mark as sensitive are redacted before they are stored.