Glossary

AI agent observability

AI agent observability is the ability to see what an agent did, step by step, and why. It covers the calls the agent made, the content it read and sent, and the decisions taken along the way.

In a run

Each step of a run shows as it happens: allowed, refused, redacted or sent to a person. The record holds which agent acted, for whom, under which credential, what was refused and why, who approved, and when.

Our word for it

The record; you can check it.

Replay, not reconstruction

The record is written as each step happens, not assembled afterwards. See AI agent audit trail for what each line holds. For any step you can trace back to the person or event that started the job, and forward to every handoff after it.

Each kind of step carries its own detail. A refusal names what was outside permission. A redaction names the filters that fired, the session that allowed the exchange, the parties and the time. An approval names the person who decided.

A record someone else can check

Logs in a database the vendor controls ask the reader to trust the vendor. Export the Ares record as a replay pack and check it with the open verifier, offline, on a machine that is not ours.

Where it stops

The record shows what agents did, not whether a business rule was wise. It does not measure whether your model gives correct answers, and it covers only the agents you enrol. It does not measure latency, cost or model errors. Keep the observability tools you use for those.

Get the open verifier, or download a sample record from the guarantees page.

Related terms

  • AI agent audit trailAn AI agent audit trail is the record of what an agent did, for whom and under what permission, written as each step happens.
  • Replay packA replay pack is one file that holds the record of what AI agents did for one job.

Tell us about one workflow.

We will say whether an agent should do it, and what the first step costs.

A senior member of the team replies within two business days.