Supervisor, router or swarm?

A multi-agent system is several agent loops, each with its own instructions, tools, permissions and often its own owner, connected by a topology that decides who acts next and what each agent gets to see. The router classifies once and dispatches once. The supervisor (orchestrator-workers) calls workers, reads their results and decides the next step until it finishes. A hierarchy stacks supervisors. A swarm lets agents hand control to each other directly. A blackboard coordinates agents through a shared store. A pipeline runs agents in a fixed order. Each topology trades control, cost and latency differently, and each adds failure modes a single agent does not have: loops, contradictions, lost context and the question "which agent did this?". Go multi-agent when you need isolation (context, permissions, ownership) or parallelism, not because the org chart has boxes.

Key takeaways:

• The first design question is "why not one agent with these tools?" Answer it in writing.
• Every hop costs a model call, tokens and latency. Supervisors add a call per hop; parallel fan-out cuts wall-clock time but not tokens.
• Put the routing decision in code where you can (a deterministic router, a pipeline edge) and in a model only where the input really varies.
• Every topology needs a hop budget, loop detection, an owner for the final answer, and agent IDs on every trace span.
• Share typed state inside a trust boundary; pass messages with contracts across one.