What organizational patterns work best for multi-agent coordination?

Josh Clark proposes that "just like people work together in many ways, agents can adopt different organizational forms, too," and names three by analogy to human work structures. "Simple automations resemble fire brigades, a straightforward hand-off from one agent to the next." More complex processes "echo critique-driven design studios, where one agent produces work and others review it, sending it back for revision until it meets the standard." A third form is "like construction crews, swarms of agents working in parallel on different parts of the same structure, coordinated by a conductor agent as general contractor."

Clark's reasoning for the third form is explicit about the trade-off it makes: "Swarms trade the order of a queue for the speed of parallel effort." The conductor agent "breaks down a phase into slices and dispatches them to a crowd of subagents, like a team of carpenters framing a house." He scopes where this fits: swarms suit "broad information-gathering that involve pursuing many directions at once," such as "research, code generation, or data processing that can be compartmentalized into separate independent areas of work." He notes the industry is currently applying this pattern narrowly — "in support of individual tasks," citing "Claude Code and other coding platforms" as doing this "routinely to work on an individual developer's project" — rather than at the population scale his broader essay argues designers must also plan for (see How do multiple agents reshape systems they were not designed for?).

The swarm form's assumption — that a task can be "compartmentalized into separate independent areas of work" — is directly contested by Why do parallel subagents produce conflicting design decisions?, which argues that subagents given what looks like the same subtask still diverge because each acts on assumptions the other never saw, and recommends defaulting to a single linear agent rather than parallel subagents. Clark's carpenters-framing-a-house image depicts independence as achievable through good decomposition; Yan's argument is that this independence is rarely clean in practice, since context slices carry implicit decisions that are never actually isolated from each other.

Clark offers the three forms as a descriptive analogy, not a tested taxonomy: the excerpt gives no criteria for choosing among fire-brigade, critique-studio, and construction-crew forms for a given task, nor any measurement of coordination cost or failure rate for any of them. It does not establish that these are the only forms, or that they are mutually exclusive in a real system. The implication is that this is a naming vocabulary useful for describing what a multi-agent system is doing, not yet a decision framework for designing one.