Supervisor agent
In hierarchical multi-agent systems, the top-level agent that delegates work to specialist sub-agents, monitors progress, handles failures, and aggregates results.
The supervisor (or "manager" or "orchestrator") agent is what makes hierarchical multi-agent systems work. It receives the high-level goal, decides which specialist sub-agents to invoke, gives each a focused sub-task, monitors their progress, retries or escalates on failure, and combines outputs into a final result.
Supervisor agents are typically frontier-model-backed (Claude Sonnet 4.6, GPT-5) because they need strong reasoning and planning. Worker sub-agents can often run on cheaper, faster models tuned for specific tasks (code generation, web search, data extraction).
In 2026 most production multi-agent systems have a supervisor pattern under the hood. LangGraph, CrewAI, and Anthropic's Agent SDK all support supervisor-style orchestration as a first-class pattern.
Frequently asked
Does my agent need a supervisor?+
If you have multiple specialist sub-agents, yes. The supervisor coordinates them. For single-agent systems, no — the agent supervises itself.
Supervisor on frontier model, workers on cheaper models?+
Yes, that is the most cost-effective production pattern. Supervisor needs reasoning depth (frontier model). Workers handle focused tasks that smaller models do well (distilled or mid-tier models).