agent-supervisor-pattern

Agent Supervisor Pattern

Design centralized orchestration systems where a supervisor agent coordinates specialized worker agents.

When to Use

• Complex tasks requiring multiple specialized agents
• Need clear visibility into agent decision-making
• Workflows requiring sequential or conditional agent execution
• Systems where human oversight is important
• Tasks with clear subtask decomposition

Architecture Overview

                    ┌─────────────┐
                    │  SUPERVISOR │
                    │   AGENT     │
                    └──────┬──────┘
                           │
           ┌───────────────┼───────────────┐
           │               │               │
           ▼               ▼               ▼
    ┌──────────┐    ┌──────────┐    ┌──────────┐
    │ Research │    │  Writer  │    │ Reviewer │
    │  Agent   │    │  Agent   │    │  Agent   │
    └──────────┘    └──────────┘    └──────────┘

Core Components

1. Supervisor Agent

The brain that orchestrates the workflow.

interface SupervisorConfig {
  name: string;
  description: string;
  workers: WorkerAgent[];
  maxIterations: number;
  routingStrategy: 'sequential' | 'parallel' | 'conditional';
}

const supervisorPrompt = `
You are a supervisor agent coordinating specialized workers.

Available Workers:
{{#each workers}}
- {{name}}: {{description}}
{{/each}}

Your responsibilities:
1. Analyze the incoming task
2. Decompose into subtasks for appropriate workers
3. Route subtasks to the right worker
4. Synthesize results into final output
5. Handle errors and retry if needed

For each step, output:
{
  "reasoning": "Why this worker for this subtask",
  "worker": "worker_name",
  "task": "Specific instructions for the worker",
  "expectedOutput": "What you expect back"
}

When complete, output:
{
  "status": "complete",
  "result": "Final synthesized result"
}
`;

2. Worker Agents

Specialized agents for specific tasks.

interface WorkerAgent {
  name: string;
  description: string;
  systemPrompt: string;
  tools: Tool[];
  outputSchema?: JSONSchema;
}

const researchWorker: WorkerAgent = {
  name: 'researcher',
  description: 'Searches and analyzes information from various sources',
  systemPrompt: `You are a research specialist.
    Search for accurate, up-to-date information.
    Always cite sources.
    Return structured findings.`,
  tools: [webSearch, documentReader],
  outputSchema: {
    type: 'object',
    properties: {
      findings: { type: 'array' },
      sources: { type: 'array' },
      confidence: { type: 'number' }
    }
  }
};