The supervisor/worker topology assigns orchestration to one coordinator agent that routes subtasks to specialized workers, then folds their outputs into a single result. It is the hub-and-spoke answer to agent sprawl.
A flat pool of peers scales badly: every agent talking to every other agent produces O(n²) message traffic and diffuse responsibility. The supervisor pattern collapses that to O(n) — workers talk only to the coordinator, never to each other. The supervisor holds the goal, decomposes it, dispatches subtasks with explicit contracts, and integrates results. Workers stay small and sharp: one per tool surface or domain, each with a context window containing only what its subtask needs.
The quiet win is context isolation. A worker retrieving from a 400-page corpus never sees the debate about deployment strategy; its window stays clean, its attention undiluted. The supervisor absorbs the coordination overhead that would otherwise pollute every worker.
| Property | Supervisor / Worker | Flat peer mesh |
|---|---|---|
| Message complexity | O(n) — hub absorbs routing | O(n²) — every peer pairs |
| Failure behavior | Coordinator is a single point of failure | Degrades gracefully, hard to diagnose |
| Context hygiene | Workers see only their subtask | Coordination noise everywhere |
| Best fit | Decomposable goals, clear specialisms | Small collaborative debates, brainstorming |
Design Notes
- Give workers typed contracts — schemas for what they accept and return — so the supervisor can validate before integrating.
- Cap worker output size; a worker that returns everything forces the supervisor to re-do the reading it delegated.
- Add a supervisor verification step before declaring completion — the pattern's failure mode is a coordinator that forwards confident nonsense unchecked.