Multi-Agent Wiki
Control

Supervisor / Manager

A primary agent plans, routes, and synthesizes; specialists execute subtasks.

Live visualizationAnimated topology — press Space to play / pause
·Diagram
Live
Variant
1 / 8

—

Definition

A primary agent plans, routes, and synthesizes results; other agents act as specialists that execute subtasks.

Category: Control structure

When to use

Production systems, task decomposition, customer-support triage, internal R&D platforms, anywhere you need stable control and observability.

When not to use

Open-ended exploration, free-form negotiation among agents, or cases where the supervisor cannot meaningfully evaluate subtask quality.

How to implement

  1. Define a SupervisorAgent that does only intent recognition, planning, routing, and synthesis — never the heavy lifting.
  2. Each specialist has its own instructions, tool set, memory scope, and permissions.
  3. When the supervisor calls a sub-agent it passes a structured task: goal / context / constraints / expected_output.
  4. Sub-agents return structured results: status / answer / evidence / next_actions / confidence.
  5. The supervisor decides next: call again, fan out in parallel, enter a verifier, request user confirmation, or finalize.

Minimal pseudocode

TypeScript
type AgentResult = {
  status: "success" | "blocked" | "need_input" | "failed";
  answer: string;
  evidence?: string[];
  confidence?: number;
};

async function supervisor(task: UserTask) {
  const plan = await planner.run(task);
  const results = [];
  for (const step of plan.steps) {
    const agent = registry.pick(step.requiredSkill);
    results.push(await agent.run({ goal: step.goal, context: task.context }));
  }
  return synthesizer.run({ task, results });
}

Recommended trace events

  • supervisor.plan.created
  • agent.task.assigned
  • agent.result.received
  • supervisor.synthesis.completed

Common failure modes

  • The supervisor becomes a bottleneck; every error concentrates on one node.
  • Sub-agent outputs are not verifiable; the supervisor blindly trusts them.
  • Full context is dumped into the supervisor, exploding tokens and erasing permission boundaries.

Implementation checklist

  • Input/output schemas defined.
  • Each agent's permission boundary defined.
  • Every agent call carries a run id / trace id.
  • Failure, timeout, cancel, and retry strategies defined.
  • Context passed is the minimum required, not the full history.
  • High-risk actions are gated by approval or a verifier.

References