What is a supervisor agent?
Updated November 2026
A supervisor agent is an agent that delegates work to specialist agents and evaluates or combines their results. Unlike a general workflow controller, it uses model-based judgment for at least part of task assignment, review, or synthesis.
How does supervised delegation work?
In LangChain’s subagent pattern, a main agent invokes subagents as tools and keeps routing centralized. The supervisor can select a specialist, supply a bounded assignment, and decide whether the returned result answers the original request. Specialists need not share the supervisor’s entire conversation.
Give each assignment an expected output format, relevant evidence, and a completion criterion. A specialist asked to identify contract renewal dates should return dates with source references, not an unconstrained essay that the supervisor must interpret again.
How is a supervisor different from an orchestrator?
CrewAI’s hierarchical process uses a manager model or custom manager agent for coordination. An orchestrator is the broader execution mechanism; a supervisor supplies a particular form of decision making within it. Keep infrastructure permissions separate from the model’s ability to recommend a next step.
| Responsibility | Suitable owner |
|---|---|
| Choose a specialist | Supervisor judgment or routing rule |
| Enforce a timeout | Workflow runtime |
| Authorize a sensitive write | Policy and approval control |
Why can supervision fail?
A supervisor can misroute requests, overlook contradictory evidence, or accept a fluent but unsupported answer. Adding a reviewer agent does not make review independent when both agents share the same mistaken assumptions or missing sources.
Central review also consumes context and adds latency. Return compact structured findings with evidence references, and reserve detailed artifacts for inspection when needed. Put limits on repeated delegation so a disagreement does not create an endless request for revisions.
For high-impact decisions, use objective validation where possible: schema checks, calculations, source matching, and permission checks. Escalate unresolved uncertainty instead of letting the supervisor convert disagreement into unwarranted confidence. Measure final task correctness separately from whether every specialist responded.
What else should you know?
Must specialists report to one supervisor?
No. Central supervision is a design choice, not a universal requirement.
Can the supervisor use tools?
Yes, subject to the same permissions and validation as other agents.
Is review a correctness guarantee?
No. Review can repeat or amplify the original error.
When is a fixed workflow preferable?
Use one when routing and acceptance criteria are already explicit.