A single agent has limits. Multi-agent systems have a different set of limits. The choice of topology is an architectural commitment with downstream consequences for cost, debuggability, and accountability.
Top-down
Fits when
A clear decomposition exists. The supervisor acts as the planner and accountable owner; sub-agents are specialists.
Structural cost
Single point of failure. The supervisor's reasoning bottlenecks the whole workflow. Easiest to audit.
Lateral
Fits when
Tasks are emergent, parallelizable, and tolerant of duplicate work. Useful for exploration and divergent thinking.
Structural cost
Non-determinism. Two runs of the same input produce different traces. Hardest to audit; expect chatty token bills.
Mediated
Fits when
Specialist agents need to coordinate without direct communication — pattern fits long-running, multi-stage workflows.
Structural cost
The shared state becomes the integration contract. Schema drift breaks every agent. Locking and consistency are now your problem.