AT-201a · Module 1

Team Topology Patterns

4 min read

There are four fundamental topologies for multi-agent teams. Hub-and-spoke, pipeline, fan-out, and mesh. Each is optimized for a different type of workload, and choosing the wrong topology is the second most common architectural mistake I see — right behind undefined roles.

Hub-and-spoke: one coordinator at the center, specialists radiating outward. All communication flows through the hub. This is the default topology for most teams because it is the simplest to reason about and the easiest to debug. The coordinator dispatches tasks, collects results, and synthesizes. No specialist talks to another specialist directly. If you are building your first agent team, start here.

Pipeline: sequential processing where each agent's output feeds the next agent's input. Research produces data. Drafting consumes data and produces prose. Review consumes prose and produces feedback. This topology is right when the work has a natural order of operations and each stage fundamentally transforms the data. Content production, report generation, and code review are all pipeline workloads.

Fan-out: one coordinator dispatches the same task template to N agents working in parallel on different inputs. Five research agents each investigating a different competitor. Four analysts each reviewing a different market segment. Fan-out maximizes throughput for independent, parallelizable work. The coordinator collects all results and synthesizes.

Mesh: agents communicate directly with each other, not just through a coordinator. This is the most complex topology and the hardest to debug. Use it only when agents need to iterate on shared artifacts — debate patterns, collaborative writing, consensus-building. Mesh topologies are powerful but expensive and should be reserved for high-stakes decisions where the cost of inter-agent communication is justified by the quality improvement.

Do This

  • Start with hub-and-spoke — it is the simplest to build and debug
  • Use pipeline when each stage transforms the data in a meaningful way
  • Use fan-out when you have N independent instances of the same task
  • Reserve mesh for debate and consensus patterns where inter-agent communication adds value

Avoid This

  • Choose mesh because it sounds sophisticated — it is expensive and hard to debug
  • Use pipeline for independent tasks that could run in parallel — that wastes time
  • Let fan-out agents depend on each other's output — that is a pipeline, not a fan-out
  • Mix topologies without clear boundaries — decide which topology governs each phase of work