Supervisor Agent

Agentic AIArchitecture and orchestrationPublished By Simon Budziak

A supervisor agent is an AI agent that assigns work to specialized worker agents, passes them the relevant context, checks their results, and decides what should happen next. It coordinates the workflow without trying to perform every task itself, which keeps each worker focused on one bounded job.

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.

Frequently asked questions

Is a supervisor agent the same as an orchestrator?

Often, yes. Both coordinate workers and decide the next step. A supervisor usually implies an agent that makes routing decisions from the current state, while an orchestrator can also be deterministic workflow code.

When does a supervisor agent help?

It helps when distinct workers need different tools, prompts, permissions, or context. A single agent is usually simpler when one worker can safely finish the task.

Summarize this page with

See this working in a system we built