Supervisor Pattern

The supervisor pattern is the most common multi-agent architecture in production. A single supervisor agent receives the user's request, decomposes it if needed, and routes each piece to the most appropriate specialist worker (for example: a researcher, a writer, a coder, a fact-checker). Workers do not talk to each other. They report back to the supervisor, which decides what to do with the results: synthesize them, dispatch follow-up work, or return the final answer. LangGraph documents this as the canonical multi-agent pattern, and OpenAI's Agents SDK uses the same model under the name "handoffs."

Two failure modes are characteristic. First, supervisor bottleneck: every interaction passes through the supervisor, which becomes a context-window and latency hotspot. Mitigation: keep the supervisor stateless when possible, summarize worker outputs before re-injecting. Second, supervisor confusion: with too many workers (more than about 5 to 7), the supervisor struggles to pick the right one consistently. Mitigation: cluster workers into teams with their own sub-supervisors, creating a hierarchy. The supervisor pattern works because it preserves a clear chain of responsibility and a single audit log, both of which become critical when an agent does the wrong thing in production.

Supervisor pattern. One agent at the top dispatches to specialists. Workers report back, never talk to each other. Used in LangGraph and OpenAI Agents SDK.

Use the supervisor pattern when one task needs more than one specialist toolset (research + writing + verification, for example), when you need a single audit point, or when different parts of the work have different latency or cost profiles that justify isolating them.

• LangGraph: Supervisor multi-agent pattern
• OpenAI Agents SDK: Handoffs
• Anthropic: Building Effective Agents

Related Terms