Every Cairn run produces a trace: an ordered record of what triggered the run, what the agent read, which tools it called, what it decided and what changed in the world as a result.

Steps, not tokens

We model a run as typed steps — trigger, research, decision, action, result — rather than a raw transcript. Steps are what humans reason about, and they map cleanly to permissions and approvals.

Tool calls are first-class

Each tool call records inputs, outputs, latency, retries and the idempotency key used. When something goes wrong, you can see whether the agent made a bad choice or the API returned something unexpected.

• Inputs and outputs, with PII redaction applied

• Duration, retries and error codes

• The policy version the decision referenced

• Who approved the step, if approval was required

Replay

Because tool outputs are recorded, any run can be replayed deterministically against a new version of the agent. That’s how evals work in Cairn: last month’s real runs become this month’s regression suite.