What is an Agent Decision Record

The category Tenbri belongs to, the properties a record must have to count as one, and the Tenbri feature behind each.

An Agent Decision Record is a signed, append-only record of every decision an AI agent makes: who acted, under which rule and inputs, who approved, and what happened, verifiable offline by anyone. Tenbri is an open-source implementation.

It is not a log or a trace. Logs and observability tools such as Langfuse and MLflow show how an agent ran, for the people who operate it. A decision record proves what the agent decided and on whose authority, to people who do not have to trust the operator. See Why not just logs?.

The minimum properties

A record counts as an Agent Decision Record when it has all of these:

PropertyWhat it meansIn Tenbri
Intent before the actionThe decision is written down before the side effect runs, so a crash cannot hide itA decision entry with phase intent, written by @tenbri.decision, tenbri append or POST /v1/streams/{stream}/intents before the call
Outcome afterWhat happened is recorded and tied to its intentA decision entry with phase outcome and status (ok, refused, error, unknown), linked by intent
PrincipalThe specific agent, person or service that acted, not a shared accountprincipal {kind, id}, optionally with its own attestation signature. See Keys
Rule and input versionsWhich rule allowed or refused it, and exactly what the agent sawrule {ref, version, result} and inputs pinned by reference, version and keyed digest; tenbri query --rule
ApprovalA person's sign-off is its own signed fact, not a chat messageAn approval entry, referenced from the decision by approval.entry; schemas can require one
Tamper evidenceAny edit, deletion, reordering or fork is detectableCanonical bytes, a hash chain, Ed25519 signatures, signed HEAD and checkpoints. See Integrity
Offline verificationAnyone can check it with no access to your systemstenbri verify against a pinned fingerprint, and receipts that prove single entries
Erasure without breaking the chainPersonal data can be deleted while the record still verifiesOnly hashes in the chain, encrypted bodies and per-subject keys; tenbri erase. See Privacy

A system missing any one of these is still useful, but it cannot answer the question a dispute, incident or audit asks: what did the agent decide, and on whose authority?

Next

Edit on GitHub

On this page