Skip to main content
A Trace is the recording of a single agent run. Every model call, every tool invocation, every retrieval, every memory op — captured with inputs, outputs, timing, and cost. Traces are how you understand what your Agents actually did in production. Not sampled — every run.

​
When you’ll look at a Trace

  • A user reported a bad answer.
  • Costs jumped and you want to know why.
  • Latency spiked and you’re hunting the slow step.
  • You’re validating a new version and want to spot-check runs.
  • You’re building a dataset from real production behavior.

​
What’s in a Trace

Inspect a run

The main view — timeline of everything that happened.

Tool calls

What each tool was called with, what it returned.

Grounding

Which parts of the answer are backed by retrieved evidence.

Cost & latency

Per-step cost and time.

Feedback

Add corrections and thumbs from the Trace itself.

Search & filter

Find runs by any criterion.

​
Retention

Traces are retained per your Flow settings (default 90 days). Aggregate metrics are retained longer. Payloads (prompts, tool args) can be scrubbed sooner than the Trace itself for compliance workloads.

​
Access

By default all workspace members can view Traces. Restrict via Flow access settings — you can allow “view Trace metadata” without allowing “view payloads” for teammates who need aggregate stats but not raw content.

​
Sharing a Trace

Every Trace has a permalink. Share with a teammate; they see the same view. External sharing (public link) is off by default and only enable-able by Owners.