What should an AI agent log during a long task?

Logs get written to satisfy a general sense that there should be a record, and they end up recording the wrong thing: a stream of activity. That format answers what happened, which is rarely the question. The questions actually asked after an agent run are why did it choose this, did the thing it says it did actually happen, and what did it not do. A log should be designed backwards from those three.

Decisions with reasons come first. Whenever the agent chose between approaches, skipped something, interpreted an ambiguous instruction, or concluded that something was out of scope, that is a decision and the reason is the useful half. This is also the content that lets a later session — or a person — tell whether the decision still holds, which a bare record of the choice cannot support.

Actions with observed results come second, and the emphasis is on observed. A log entry saying a record was updated is worth much less than one saying the update was issued and the record now reads a particular way. The gap between the two is where the most common false completion lives: the command was sent, something went wrong, nothing checked, and the log records the intention as though it were an outcome.

Failures and refusals come third and are the most frequently omitted. What could not be read, what returned an error, what was attempted and abandoned, what was ambiguous. These entries feel like admissions of trouble and they are the highest-value content in the file, because they are the only place the boundary of the work is visible. A log showing an unbroken sequence of successes on a long job is describing a run nobody has looked at closely.

There is also a counting layer that is worth keeping separate from prose. How many items existed, how many were processed, how many skipped, how long it took, what it cost. These are the entries that make it possible to compare runs, notice drift, and answer questions about whether the thing is still working. Buried in narrative they are unusable; kept as a small structured summary they are the first thing anyone reads.

Two things to leave out. Restating the plan, which is already written down and only creates a second version that can disagree with the first. And a general commentary on progress, which is the bulk of most agent logs and is the content nobody has ever gone back to read. The test for any candidate log line is whether it would change what somebody does when something turns out to be wrong; most narration would not.