Multi-Agent Orchestration
Use multiple agents when distinct responsibilities need different context, tools or evaluation criteria. Adding roles does not automatically improve quality; it also adds handoffs, latency and opportunities to lose evidence.
#Three orchestration patterns
| Pattern | Suitable task | Main risk |
|---|---|---|
| Sequential | Research, then write, then review | A weak handoff contaminates every later step |
| Hierarchical | A coordinator delegates bounded subtasks | Coordinator becomes a bottleneck or duplicates work |
| Conversational | Roles critique or negotiate alternatives | Repeated discussion without a stopping rule |
CrewAI illustrates role/task-oriented coordination. AutoGen illustrates agent interaction patterns. Their APIs and supported integrations change; use documentation for a pinned release rather than copying an old import blindly.
#Define the handoff contract
A researcher should return claims, source URLs, publication dates, uncertainty and unresolved questions. A writer should use that packet rather than inventing missing evidence. A reviewer checks the resulting artifact against the packet and acceptance criteria.
Give each role only its required capabilities. For example, a writer can receive source extracts without gaining permission to browse private databases or publish externally. Shared state should preserve source IDs so facts remain traceable across roles.
#Bound the team
Set limits for turns, retries, parallel work and total usage. Decide which role owns the final deliverable and which conditions require a stop. Independent tasks can run concurrently; dependent work must wait for its input.
Compare the team with a single-agent or fixed-workflow baseline on the same cases. Measure factual accuracy and completed outcomes, not number of messages. Keep a smaller design if the extra roles add no measurable benefit.
