Multi-agent orchestration with supervisor agents, the pattern that finally worked for us
by•
Tried flat multi-agent setups, didn't scale. Tried hierarchical with strict roles, too rigid. Settled on supervisor pattern with dynamic delegation.
How it works:
- Supervisor agent receives the task, decomposes it
- Delegates subtasks to specialised worker agents (research, write, review)
- Worker agents return structured output
- Supervisor synthesises and either ships or re-delegates
- Human review gate at the synthesis step for critical tasks
What made it scale:
- Worker agents are stateless, supervisor holds state
- Strict input/output schemas per agent
- Cost ceiling at the supervisor level prevents runaway loops
What's the multi-agent pattern that's working in your production setup?
3 views
Replies
The supervisor-holds-state + stateless-workers split is the right call within a run — the place I'd watch is across runs. When all the state lives in the supervisor's context window, the judgment the system built up (what a worker already tried, which delegation paths dead-ended, prior decisions) evaporates when the run ends — so every task re-derives it.
Externalizing that state to a durable store — the supervisor reloads prior decisions with their provenance instead of rebuilding context each run, and the stateless workers can read from it too — is what turns "works per run" into "compounds over time."
The cost ceiling at the supervisor level is a sharp touch. Runaway delegation loops are the quiet killer.