Supervisor Agents

Introduction

A supervisor agent assigns tasks to specialists, evaluates their output and decides next steps. It is the brain of the multi-agent system.

Beginner analogy: the team lead in a sprint — assigns tickets, reviews PRs, escalates.

Understanding the topic

Core concepts:

• Holds the goal and the plan.
• Routes tasks to the right specialist.
• Aggregates and merges results.
• Detects loops and escalates.
• Often a stronger model than specialists.

Syntax reference

Visual workflow / architecture:

      ┌────────────┐
      │ Supervisor │
      └──┬─────┬───┘
         │     │
         ▼     ▼
       Spec1  Spec2  …

Real-world use

LangGraph 'supervisor' template, CrewAI hierarchical crews, OpenAI Swarm hand-off agents.

Best practices

• Use a stronger LLM for the supervisor.
• Limit specialist count (≤ 5 typically).

Common mistakes

• Supervisor that micromanages — defeats specialisation.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Three jobs of a supervisor.
• Q2. Why often a stronger model?
• Q3. Scenario: supervisor keeps re-routing to the wrong specialist. Two fixes?