The audit trail

For agents doing legal and financial work, “what did it do, and why?” isn’t a nice-to-have — it’s a requirement. KAOS agents keep an audit trail: a durable, secondary record of the turn, captured from the same event stream that drives the live response.

A separate data plane

The audit trail is not the response and not the memory — it’s a third thing, recorded alongside them. As a turn runs, a recorder writes its events (tool calls, steps, citations, usage, refusals, errors) to a structured, append-only log. Because it’s derived from the typed event stream, it captures the reasoning shape of the turn, not just the final text.

The kaos-agents examples ship an HTML viewer that renders these JSONL audit records — you can replay exactly what the agent did, step by step.

Built for trust, with redaction

Two properties make the audit trail safe to keep:

• Redaction. Secrets, credentials, and sensitive payloads are kept out of the record — the trail is designed to be retained and reviewed without leaking what shouldn’t be in it.
• Honest coverage. Not every internal is captured, and the docs are candid about the gaps rather than implying total coverage. A trail you can trust is one whose limits are documented.

Why it matters

• Accountability. You can answer “why did the agent cite this / call that tool / refuse here?” after the fact.
• Debugging. A failed or surprising turn can be replayed from its record instead of reproduced live.
• Review. Compliance and security reviewers get an evidence trail without having to instrument the agent themselves.

This is the same stance as the refusal contract and cost as a contract: make the trustworthy thing structural, observable, and honest about its limits.