When you have one agent, you have an agent. When you have ten, you have a topology problem. The shape of that topology determines almost everything about how the system evolves: who can change what, where bugs hide, what scales linearly vs. what scales catastrophically.
Three topologies
- Centralized — one router agent dispatches to specialist agents. Easy to govern, easy to bottleneck. Use when consistency matters more than autonomy.
- Federated — domain teams own their agents, share a thin set of platform primitives (eval, retrieval, identity). Use when domain expertise is the scarce resource.
- Peer-to-peer — agents call each other directly with no router. Powerful and brittle in equal measure. Use sparingly, with strong observability, or you will lose a Tuesday.
The choice maps to your org chart
Conway's Law applies. If your org is one centralized team, build centralized. If you have strong domain teams already, federated is the path of least resistance. If you skip ahead to peer-to-peer without the observability to back it, you'll spend a quarter learning why.
- 01Anthropic — Multi-agent research system ↗
A worked example of orchestrator + worker topology.
Knowledge check
0/1 answered1. Your org has strong domain teams (sales, support, ops) each already shipping their own AI features. The natural topology is:
Discussion
0 commentsBe the first to start the conversation.