LangGraph describes its supervisor pattern as a central agent that controls delegation and communication among specialized workers. That is the useful distinction: a supervisor owns the route, while workers own the task they were given.
How does a supervisor agent coordinate workers?
The supervisor reads shared state, chooses a worker, sends a bounded task, then receives an outcome before selecting the next route. It should pass only the context the next worker needs, rather than every message and tool result. A subagent can research, write, or call a dedicated system while the supervisor keeps the goal and stop condition intact.
This pattern is one form of a multi-agent system. It resembles the orchestrator-worker pattern, though an orchestrator can use fixed routing rules instead of an LLM to choose the worker.
When is a supervisor the wrong choice?
A supervisor adds another model call, more state, and another place for routing to fail. Use it only when workers genuinely need separate tools, permissions, or expertise. A single AI agent is easier to evaluate when one loop can complete the job. When a worker returns control, an explicit agent handoff keeps ownership of the next decision clear.