How agents talk determines what the fleet can know, how it fails, and what it costs. Mesh, star, hierarchy, and pipeline each encode a trade-off between shared context and coordination overhead — and the wrong topology is fatal at scale in ways no prompt can fix.

Mesh — everyone talks to everyone — maximizes shared information and minimizes bottlenecks, at O(n²) message cost and diffuse accountability; it works for debates and small creative swarms, and collapses by a dozen participants. Star — a central hub relaying all traffic — makes costs linear and accountability singular, at the price of a hot spot whose context window becomes the system's real constraint. Hierarchy — teams of agents led by lead agents, led by a chief — matches human organizations because it solves the same problem: spanning trees keep every conversation local while the org chart carries summaries upward and decisions downward. Pipeline — a strict producer-consumer chain — is what plan-and-execute or staged generation really is, trading flexibility for total auditability.

The research on structured topologies (AutoGen's programmable conversation patterns, agent workflow papers through 2024–25) consistently lands on the same conclusions: constraining who may speak to whom improves both quality and cost, hierarchical structures scale furthest, and free-form group chat is where multi-agent systems go to burn tokens agreeing with each other.

Choosing by Question

• Do participants need each other's full reasoning? Yes → small mesh (debate, critique). No → star or hierarchy with summarized edges.
• Is the work decomposable in advance? Yes → pipeline: stages, contracts, no chatter.
• Does the hub's context saturate? That is the star's ceiling — move to hierarchy before, not after, the coordinator degrades.
• What must survive an audit? Topology is the audit structure: a pipeline is a paper trail; a mesh is a mystery.

The deepest habit to steal from distributed systems: treat communication as the expensive thing it is. Message passing — between humans, processes, or agents — is where complexity lives. The topology is not an implementation detail of the fleet; it is the fleet's actual architecture.