Day 3: Multi-Agent Systems - The Supervisor Pattern

Source: Dev.to

Introduction

Welcome to Day 3 of the 4‑Day Series Agentic AI with LangChain/LangGraph. When tasks become complex, a single agent that tries to be a “Researcher, Writer, Editor, and Coder” quickly gets overwhelmed. The solution is a Multi‑Agent System that splits the brain into specialized personas that collaborate via a shared state.

Overview of the Supervisor Pattern

We’ll build a simple graph where two agents work together:

• Researcher – has access to search tools; system prompt: “You are a researcher…”.
• Writer – has no tools; system prompt: “You are a writer…”.

The agents communicate by appending messages to a shared state. Only the Researcher is bound to tools, preventing the Writer from accidentally invoking them.

Creating Agents

// Helper to create an agent node
const createAgent = (model, systemPrompt, tools = []) => {
  // Bind tools only to the agents that need them
  const modelWithTools = model.bindTools(tools);
  return async (state) => {
    // Prepend the system prompt to the conversation history
    const messages = [
      new SystemMessage(systemPrompt),
      ...state.messages,
    ];
    const response = await modelWithTools.invoke(messages);
    // Return the new message(s) to be added to the shared state
    return { messages: [response] };
  };
};

The Researcher receives the tools array (e.g., a search tool), while the Writer receives an empty array.

Deterministic Supervisor Flow

For this tutorial we use a fixed sequence instead of a dynamic “Supervisor” LLM:

Researcher → (maybe Tools) → Researcher → Writer → End

graph TD
    Start[__start__] --> Researcher
    Researcher -->|Call Tool?| Condition{Has Tool Call?}
    Condition -->|Yes| Tools[ToolNode]
    Condition -->|No| Writer
    Tools --> Researcher
    Writer --> End[__end__]

The flow is encoded in LangGraph edges (see next section).

Defining the Workflow in LangGraph

const workflow = new StateGraph(MessagesAnnotation)
  .addNode("researcher", researcherNode)
  .addNode("tools", toolNode)
  .addNode("writer", writerNode)

  // Start the graph with the Researcher
  .addEdge("__start__", "researcher")

  // Researcher logic:
  // – If it calls a tool → go to "tools"
  // – If it produces an answer → hand off to "writer"
  .addConditionalEdges("researcher", (state) => {
    const last = state.messages[state.messages.length - 1];
    return last.tool_calls?.length ? "tools" : "writer";
  })

  // Tools always return results back to the Researcher
  .addEdge("tools", "researcher")

  // Writer always finishes the job
  .addEdge("writer", "__end__");

How Messages Are Passed

1. Researcher adds a message such as “I found X, Y, Z.” to the shared state.
2. The graph transitions to Writer.
3. Writer receives the full conversation (UserMessage, ResearcherMessage, …) and generates the final output, e.g., “Here is a blog post about X, Y, Z…”.

The Writer never needs to perform a search; it simply “reads” what the Researcher wrote. This separation of concerns lets you tune each agent independently (e.g., temperature 0.8 for creative writing, temperature 0 for factual research).