Agent audit trails and replay: explaining decisions safely

An activity log is not yet an explanation

A list of timestamps and model responses can show that work occurred while leaving the important questions unanswered. Who was allowed to decide? Which earlier record justified the transition? What actually changed outside the runtime?

The audit reference identifies actor, scope, predecessor, policy context, decision, effect and outcome as relevant evidence. Use those relationships to structure an investigation rather than treating a large volume of logs as proof of accountability.

Retain references that connect the decision

In an illustrative purchase proposal, retain the reviewed artifact reference, the specific approval, the assignment context and the effect receipt. An investigator should be able to follow the decision into the action without guessing from similar text in a conversation.

The evidence overview covers membership changes, assignments, recovery signals and outcomes as well as actions. Those surrounding transitions matter when an apparently valid approval was later superseded or the executor changed.

Replay should reconstruct before it executes

The replay and checkpoint guidance calls for reconstructing decisions and evidence rather than blindly repeating side effects. Separate an investigation mode from any operation that can invoke a real connector.

For example, replaying the rationale for a notification should not send the notification again. An application can inspect the original input references and retained outcome while keeping the external sink disabled. If a new effect is requested, it needs a new deliberate authorization path appropriate to the operation.

Redaction belongs in the evidence design

The audit guidance calls for recursive redaction and evidence boundaries. A trace that copies entire prompts, credentials or customer documents into coordination records can create a separate exposure from the original application.

Define which fields the operator needs to explain the transition, which payloads remain in scoped storage and which references can safely travel. Test nested data and classified errors rather than checking only a top-level log message. Retention and access policies should apply to the evidence store as deliberately as to the work itself.

Test an investigation after restart

• Choose one action and follow its actor and policy context.
• Recover its predecessor, assignment and outcome references.
• Restart the relevant storage-backed process.
• Repeat the inspection without invoking the external effect.
• Verify that sensitive nested fields remain redacted.

The PostgreSQL operations guide connects backups and retention to checkpoints and evidence. Successful persistence helps preserve the explanation; it does not make a historically invalid authority decision valid.