Pre-execution control and evidence design.
AI AGENT AUDIT TRAILS / PUBLIC EXPLANATION
What should an AI agent action log record?
An AI agent's chat transcript is not an adequate action log. The Japanese source proposes separating requests, decision inputs, approval, external execution, reason codes and timestamps so each operation can later be examined.
Do not confuse an action log with chat history
A saved conversation rarely proves who allowed a specific operation, under which conditions, and what external action actually occurred. The proposed record distinguishes inputs, targets, authority, approval and execution outcome.
- Design aim: review and reconstruct the action at the level of a specific operation.
Record stops as well as approvals
Reasons for allowing or blocking an operation should be recorded with the relevant approval reference. Reason codes make HOLD and ESCALATE traceable protective states, rather than unexplained errors.
- A log is useful only if the decision and its approval boundary can be traced.
Mark the external-action boundary
Sending email, placing an order or deploying to production can affect external systems. Do not treat these effects like an internal summary or draft. Put an auditable observation point immediately before the external step.
- Proposed evidence of an action is distinct from the action being authorized.
Related sources and materials
PoC-stage HOLD / ALLOW / DENY interface.
Japanese source on why a request passed or stopped.
Observation of public AI-reference quality; not an action-log verifier.
The approval boundary discussed in the companion answer.
doc_id: C3-WEB-ANSWERS-AI-AGENT-AUDIT-LOGS-EN-0.1 / version: 0.1.0 / status: REVIEW_REQUIRED / translated: 2026-10-13 / source_doc_id: C3-WEB-ANSWERS-AI-AGENT-AUDIT-LOGS-0.1 / source_last_updated: 2026-06-05