Cumulus · experiment 191

Coordination lab

When is it worth waiting for a batch? Give the same workers three delivery policies and watch what happens.

All durations are simulated minutes. Change a value, then compare. No model calls or wall-clock waiting.

Trade-offs for the full run

Delivery policy Meetings Coordinator min Blocked worker-min Longest wait All done at

Blocked worker-minutes add up across workers. Coordinator time includes meeting setup and answering; initial unrelated work is excluded. No score hides the trade-off.

Replay

Working Waiting for answer Meeting Initial mainline work. Hover a segment for its times.

    What this toy assumes

    Workers start together. Worker 1 asks after “work before question”; each later worker's first question is delayed by the stagger. Each answer unlocks another work period; after the last question that work finishes the task. Every question blocks its worker. Status updates are free and do not call the coordinator.

    A meeting costs setup plus time per question. Answers go to that meeting's participants together, with no acknowledgment turns. Arrivals during a meeting stay queued for the next. A busy mainline is never interrupted. At an exact tie, arrivals are included before a meeting starts.

    Activity-aware delivery waits only if at least two workers are working, have an unanswered question, or finished work within the lookback. Never-started workers do not count. The count is sampled when the first queued question arrives; the deadline stays fixed even if activity changes. This deliberately exposes stale-activity effects.

    This is a timing model, not evidence about LLM speed or decision quality. It does not model nonblocking questions, failures, urgency, dependent workers, acknowledgment storms, or the mainline's continuing task throughput. Adjust assumptions before drawing conclusions.