The Supervisor Pattern: Multi-Agent Orchestration

What happens when your problem is too big for even a "Planner"? What if you need a specialized Coder, a specialized Writer, and a specialized Security Expert?

You use the Supervisor Pattern.

1. The Managerial Brain

In this pattern, there is one "Supervisor" agent that acts as the router. It does not do the work itself. Instead, it decides which "Worker Agent" should handle the current turn.

• Supervisor: "We have a request to build a secure web app. Coder, start with the boilerplate. Security, review his code. Writer, document the process."

---

2. The Communication Loop

Unlike the Planner–Executor, where the Planner lists everything up front, the Supervisor handles Dynamic Delegation.

1. User: "Build me a data dashboard."
2. Supervisor: Hands task to "Data Specialist."
3. Data Specialist: Returns a cleaned CSV.
4. Supervisor: Views the CSV. Hands it to "Chart Specialist."
5. Chart Specialist: Returns a PNG chart.
6. Supervisor: Hands it to "Quality Checker."
7. Quality Checker: "This chart is missing a legend."
8. Supervisor: Hands it back to "Chart Specialist" with the fix instructions.

---

3. Visualizing the Orchestration

graph TD
    User[Complex Multi-Step Request] --> Sup[Supervisor Agent]
    Sup --> W1[Worker A: Researcher]
    Sup --> W2[Worker B: Programmer]
    Sup --> W3[Worker C: Technical Writer]
    W1 --> Sup
    W2 --> Sup
    W3 --> Sup
    Sup --> Final[Final Delivery]

---

4. Why This Pattern Wins

A. Dedicated Personas (System Prompts)

Each Worker has a specialized system prompt.

• The Coder is told to be "Brief and precise with syntax."
• The Writer is told to be "Creative and professional with tone." By separating them, you get higher quality than asking one model to "Be everything at once."

B. Scalability

You can add a 4th or 5th agent (e.g., a "Legal Agent") without changing the logic of the Coder. You only have to tell the Supervisor that a new expert is available.

---

5. Code Example: Supervisor logic

workers = ["Researcher", "Coder", "Writer"]

# The Supervisor prompt
prompt = f"Given this request: '{user_request}', which of these workers should take the next action: {workers}? Respond with ONLY the name."

next_worker = llm.call(prompt)

if next_worker == "Researcher":
    # Transfer control to the Researcher agent
    result = researcher_agent.run(state)
elif next_worker == "Coder":
    # Transfer control to the Coder agent
    result = coder_agent.run(state)

---

Key Takeaways

• The Supervisor is a router and manager, not a doer.
• Specialist Agents produce higher quality output because their prompts are focused.
• Communication is mediated through a Central State that the Supervisor monitors.
• This pattern mimics Human Teams (PM, Dev, QA).