Slightly fewer restrictions on final iteration

[heurihermilab]

Everyone’s mileage may vary, but one way I’ve found to manage directing agents, with their frequent memory lapses, through complex actions is to create a plan document that they can read, as well as a status document that keeps track of where they are in the plan, next steps, etc. I’ve had real success in multi-turn refactors by asking them to write current status to the status document when they receive the last-iteration nudge, and feeding that in as part of next turn’s input.

However, recently the agents are balking at writing the status document, as a result of a rule forbidding tool use during the final iteration. The limitation seems reasonable — I imagine such tool use could subvert the spirit and probably the intent of the iteration limit — but I wonder if an exception could be added specifically for writing of a status report or similar during the final iteration.