What should an AI agent log? Activity is not the same as evidence

The concrete detail comes first, the conclusion second. Below we show the reasoning, its limits and what this story can genuinely add to your own work.

One hundred “working on it” updates do not prove that a task advanced. A useful agent log lets another person reconstruct a decision and locate the point where the result stopped being reliable.

The minimum record

• a concrete goal and completion criterion,
• versions of relevant inputs,
• tools used and state changes made,
• test outcome with a timestamp,
• missing data and assumptions,
• cost or time when it affected a choice,
• the point at which human approval became necessary.

Secrets, complete personal data and indiscriminate logs do not belong here. Observability also needs a privacy boundary.

The key rule is simple: status should describe the artefact, not the agent’s confidence. “File created” and “product verified” are different states.