An audit trail for AI agents: what to record and how to prove it
When an agent acts for you, someone will later ask what it did and who allowed it. An audit trail answers that. A log the same system can rewrite does not.
What to record
- Which agent, on which connection.
- Which tool, and its risk class.
- The decision: ran, waited, approved, declined, refused, expired.
- Who decided, when, and with what proof (a session, a passkey, two people).
- A fingerprint of what was asked and of the result.
- Changes to the rules themselves: a level changed, a rule added, a person removed.
What not to record
Do not copy the contents of every request into the log. A log full of emails and file contents becomes the most sensitive store you have, and cannot be kept for long. Record a digest: it proves what was asked without holding it.
Making it verifiable
A record is evidence only if it cannot be changed quietly. Three properties matter: each entry is signed, entries are chained so one cannot be removed without breaking the next, and the system that writes the log cannot update or delete it.
Anyone holding an entry should be able to check it without trusting the system that made it.
How Operon OS Agenaut records it
Every decision Agenaut makes, and every decision you make, produces a signed receipt in your ledger. A receipt holds codes and digests, not what was asked. You can check any receipt in the app, see each step of the check, and download the ledger as a file. What was actually asked stays on the request for a limited time (90 days by default) and is then deleted, while the receipt remains.
Try it. Operon OS Agenaut is in beta. Request access, or read the docs to see how an agent is connected.
Related guides
- Guardrails for LLM agents: levels, limits and an off switch
- Keeping autonomous AI agents under control
- Enterprise AI agents: approvals, roles and the two-person rule
Terms used here are explained in the glossary.