Build custom multi-agent systems

A single agent might struggle if it needs to specialize in multiple domains or manage many tools. To tackle this, you can break your agent into smaller, independent agents and composing them into a multi-agent system. In multi-agent systems, agents need to communicate between each other. They do so via handoffs — a primitive that describes which agent to hand control to and the payload to send to that agent. This guide covers the following:

implementing handoffs between agents
using handoffs and the prebuilt agent to build a custom multi-agent system

To get started with building multi-agent systems, check out LangGraph prebuilt implementations of two of the most popular multi-agent architectures — supervisor and swarm.

Handoffs

To set up communication between the agents in a multi-agent system you can use handoffs — a pattern where one agent hands off control to another. Handoffs allow you to specify:

destination: target agent to navigate to (e.g., name of the LangGraph node to go to)
payload: information to pass to that agent (e.g., state update)

Create handoffs

To implement handoffs, you can return Command objects from your agent nodes or tools:

import { tool } from "@langchain/core/tools";
import { Command, MessagesZodState } from "@langchain/langgraph";
import { z } from "zod";

function createHandoffTool({
  agentName,
  description,
}: {
  agentName: string;
  description?: string;
}) {
  const name = `transfer_to_${agentName}`;
  const toolDescription = description || `Transfer to ${agentName}`;

  return tool(
    async (_, config) => {
      // (1)!
      const state = config.state;
      const toolCallId = config.toolCall.id;

      const toolMessage = {
        role: "tool" as const,
        content: `Successfully transferred to ${agentName}`,
        name: name,
        tool_call_id: toolCallId,
      };

      return new Command({
        // (3)!
        goto: agentName,
        // (4)!
        update: { messages: [...state.messages, toolMessage] },
        // (5)!
        graph: Command.PARENT,
      });
    },
    {
      name,
      description: toolDescription,
      schema: z.object({}),
    }
  );
}

Access the state of the agent that is calling the handoff tool through the config parameter.
The Command primitive allows specifying a state update and a node transition as a single operation, making it useful for implementing handoffs.
Name of the agent or node to hand off to.
Take the agent’s messages and add them to the parent’s state as part of the handoff. The next agent will see the parent state.
Indicate to LangGraph that we need to navigate to agent node in a parent multi-agent graph.

If you want to use tools that return Command, you can either use prebuilt create_react_agent / ToolNode components, or implement your own tool-executing node that collects Command objects returned by the tools and returns a list of them, e.g.:

const callTools = async (state) => {
  // ...
  const commands = await Promise.all(
    toolCalls.map(toolCall => toolsByName[toolCall.name].invoke(toolCall))
  );
  return commands;
};

This handoff implementation assumes that:

each agent receives overall message history (across all agents) in the multi-agent system as its input. If you want more control over agent inputs, see this section

each agent outputs its internal messages history to the overall message history of the multi-agent system. If you want more control over how agent outputs are added, wrap the agent in a separate node function:

const callHotelAssistant = async (state) => {
  // return agent's final response,
  // excluding inner monologue
  const response = await hotelAssistant.invoke(state);
  // highlight-next-line
  return { messages: [response.messages.at(-1)] };
};

Control agent inputs

You can use the Send() primitive to directly send data to the worker agents during the handoff. For example, you can request that the calling agent populate a task description for the next agent:

import { tool } from "@langchain/core/tools";
import { Command, Send, MessagesZodState } from "@langchain/langgraph";
import { z } from "zod";

function createTaskDescriptionHandoffTool({
  agentName,
  description,
}: {
  agentName: string;
  description?: string;
}) {
  const name = `transfer_to_${agentName}`;
  const toolDescription = description || `Ask ${agentName} for help.`;

  return tool(
    async (
      { taskDescription },
      config
    ) => {
      const state = config.state;
      
      const taskDescriptionMessage = {
        role: "user" as const,
        content: taskDescription,
      };
      const agentInput = {
        ...state,
        messages: [taskDescriptionMessage],
      };
      
      return new Command({
        // highlight-next-line
        goto: [new Send(agentName, agentInput)],
        graph: Command.PARENT,
      });
    },
    {
      name,
      description: toolDescription,
      schema: z.object({
        taskDescription: z
          .string()
          .describe(
            "Description of what the next agent should do, including all of the relevant context."
          ),
      }),
    }
  );
}

See the multi-agent supervisor example for a full example of using Send() in handoffs.

Build a multi-agent system

You can use handoffs in any agents built with LangGraph. We recommend using the prebuilt agent or ToolNode, as they natively support handoffs tools returning Command. Below is an example of how you can implement a multi-agent system for booking travel using handoffs:

import { createReactAgent } from "@langchain/langgraph/prebuilt";
import { StateGraph, START, MessagesZodState } from "@langchain/langgraph";
import { z } from "zod";

function createHandoffTool({
  agentName,
  description,
}: {
  agentName: string;
  description?: string;
}) {
  // same implementation as above
  // ...
  return new Command(/* ... */);
}

// Handoffs
const transferToHotelAssistant = createHandoffTool({
  agentName: "hotel_assistant",
});
const transferToFlightAssistant = createHandoffTool({
  agentName: "flight_assistant",
});

// Define agents
const flightAssistant = createReactAgent({
  llm: model,
  // highlight-next-line
  tools: [/* ... */, transferToHotelAssistant],
  // highlight-next-line
  name: "flight_assistant",
});

const hotelAssistant = createReactAgent({
  llm: model,
  // highlight-next-line
  tools: [/* ... */, transferToFlightAssistant],
  // highlight-next-line
  name: "hotel_assistant",
});

// Define multi-agent graph
const multiAgentGraph = new StateGraph(MessagesZodState)
  // highlight-next-line
  .addNode("flight_assistant", flightAssistant)
  // highlight-next-line
  .addNode("hotel_assistant", hotelAssistant)
  .addEdge(START, "flight_assistant")
  .compile();

Full example: Multi-agent system for booking travel

import { tool } from "@langchain/core/tools";
import { createReactAgent } from "@langchain/langgraph/prebuilt";
import { StateGraph, START, MessagesZodState, Command } from "@langchain/langgraph";
import { ChatAnthropic } from "@langchain/anthropic";
import { isBaseMessage } from "@langchain/core/messages";
import { z } from "zod";

// We'll use a helper to render the streamed agent outputs nicely
const prettyPrintMessages = (update: Record<string, any>) => {
  // Handle tuple case with namespace
  if (Array.isArray(update)) {
    const [ns, updateData] = update;
    // Skip parent graph updates in the printouts
    if (ns.length === 0) {
      return;
    }

    const graphId = ns[ns.length - 1].split(":")[0];
    console.log(`Update from subgraph ${graphId}:\n`);
    update = updateData;
  }

  for (const [nodeName, updateValue] of Object.entries(update)) {
    console.log(`Update from node ${nodeName}:\n`);

    const messages = updateValue.messages || [];
    for (const message of messages) {
      if (isBaseMessage(message)) {
        const textContent =
          typeof message.content === "string"
            ? message.content
            : JSON.stringify(message.content);
        console.log(`${message.getType()}: ${textContent}`);
      }
    }
    console.log("\n");
  }
};

function createHandoffTool({
  agentName,
  description,
}: {
  agentName: string;
  description?: string;
}) {
  const name = `transfer_to_${agentName}`;
  const toolDescription = description || `Transfer to ${agentName}`;

  return tool(
    async (_, config) => {
      // highlight-next-line
      const state = config.state; // (1)!
      const toolCallId = config.toolCall.id;

      const toolMessage = {
        role: "tool" as const,
        content: `Successfully transferred to ${agentName}`,
        name: name,
        tool_call_id: toolCallId,
      };

      return new Command({
        // highlight-next-line
        goto: agentName, // (3)!
        // highlight-next-line
        update: { messages: [...state.messages, toolMessage] }, // (4)!
        // highlight-next-line
        graph: Command.PARENT, // (5)!
      });
    },
    {
      name,
      description: toolDescription,
      schema: z.object({}),
    }
  );
}

// Handoffs
const transferToHotelAssistant = createHandoffTool({
  agentName: "hotel_assistant",
  description: "Transfer user to the hotel-booking assistant.",
});

const transferToFlightAssistant = createHandoffTool({
  agentName: "flight_assistant",
  description: "Transfer user to the flight-booking assistant.",
});

// Simple agent tools
const bookHotel = tool(
  async ({ hotelName }) => {
    return `Successfully booked a stay at ${hotelName}.`;
  },
  {
    name: "book_hotel",
    description: "Book a hotel",
    schema: z.object({
      hotelName: z.string(),
    }),
  }
);

const bookFlight = tool(
  async ({ fromAirport, toAirport }) => {
    return `Successfully booked a flight from ${fromAirport} to ${toAirport}.`;
  },
  {
    name: "book_flight",
    description: "Book a flight",
    schema: z.object({
      fromAirport: z.string(),
      toAirport: z.string(),
    }),
  }
);

const model = new ChatAnthropic({
  model: "claude-3-5-sonnet-latest",
});

// Define agents
const flightAssistant = createReactAgent({
  llm: model,
  // highlight-next-line
  tools: [bookFlight, transferToHotelAssistant],
  prompt: "You are a flight booking assistant",
  // highlight-next-line
  name: "flight_assistant",
});

const hotelAssistant = createReactAgent({
  llm: model,
  // highlight-next-line
  tools: [bookHotel, transferToFlightAssistant],
  prompt: "You are a hotel booking assistant",
  // highlight-next-line
  name: "hotel_assistant",
});

// Define multi-agent graph
const multiAgentGraph = new StateGraph(MessagesZodState)
  .addNode("flight_assistant", flightAssistant)
  .addNode("hotel_assistant", hotelAssistant)
  .addEdge(START, "flight_assistant")
  .compile();

// Run the multi-agent graph
const stream = await multiAgentGraph.stream(
  {
    messages: [
      {
        role: "user",
        content: "book a flight from BOS to JFK and a stay at McKittrick Hotel",
      },
    ],
  },
  // highlight-next-line
  { subgraphs: true }
);

for await (const chunk of stream) {
  prettyPrintMessages(chunk);
}

Access agent’s state
The Command primitive allows specifying a state update and a node transition as a single operation, making it useful for implementing handoffs.
Name of the agent or node to hand off to.
Take the agent’s messages and add them to the parent’s state as part of the handoff. The next agent will see the parent state.
Indicate to LangGraph that we need to navigate to agent node in a parent multi-agent graph.

Multi-turn conversation

Users might want to engage in a multi-turn conversation with one or more agents. To build a system that can handle this, you can create a node that uses an interrupt to collect user input and routes back to the active agent. The agents can then be implemented as nodes in a graph that executes agent steps and determines the next action:

Wait for user input to continue the conversation, or
Route to another agent (or back to itself, such as in a loop) via a handoff

import { interrupt, Command } from "@langchain/langgraph";

function human(state: MessagesState): Command {
  const userInput: string = interrupt("Ready for user input.");

  // Determine the active agent
  const activeAgent = /* ... */;

  return new Command({
    update: {
      messages: [{
        role: "human",
        content: userInput,
      }]
    },
    goto: activeAgent,
  });
}

function agent(state: MessagesState): Command {
  // The condition for routing/halting can be anything, e.g. LLM tool call / structured output, etc.
  const goto = getNextAgent(/* ... */); // 'agent' / 'anotherAgent'

  if (goto) {
    return new Command({
      goto,
      update: { myStateKey: "myStateValue" }
    });
  }

  return new Command({ goto: "human" });
}

Full example: multi-agent system for travel recommendations

In this example, we will build a team of travel assistant agents that can communicate with each other via handoffs.We will create 2 agents:

travel_advisor: can help with travel destination recommendations. Can ask hotel_advisor for help.
hotel_advisor: can help with hotel recommendations. Can ask travel_advisor for help.

import { ChatAnthropic } from "@langchain/anthropic";
import { StateGraph, START, MessagesZodState, Command, interrupt, MemorySaver } from "@langchain/langgraph";
import { createReactAgent } from "@langchain/langgraph/prebuilt";
import { tool } from "@langchain/core/tools";
import { z } from "zod";

const model = new ChatAnthropic({ model: "claude-3-5-sonnet-latest" });

const MultiAgentState = MessagesZodState.extend({
  lastActiveAgent: z.string().optional(),
});

// Define travel advisor tools
const getTravelRecommendations = tool(
  async () => {
    // Placeholder implementation
    return "Based on current trends, I recommend visiting Japan, Portugal, or New Zealand.";
  },
  {
    name: "get_travel_recommendations",
    description: "Get current travel destination recommendations",
    schema: z.object({}),
  }
);

const makeHandoffTool = (agentName: string) => {
  return tool(
    async (_, config) => {
      const state = config.state;
      const toolCallId = config.toolCall.id;

      const toolMessage = {
        role: "tool" as const,
        content: `Successfully transferred to ${agentName}`,
        name: `transfer_to_${agentName}`,
        tool_call_id: toolCallId,
      };

      return new Command({
        goto: agentName,
        update: { messages: [...state.messages, toolMessage] },
        graph: Command.PARENT,
      });
    },
    {
      name: `transfer_to_${agentName}`,
      description: `Transfer to ${agentName}`,
      schema: z.object({}),
    }
  );
};

const travelAdvisorTools = [
  getTravelRecommendations,
  makeHandoffTool("hotel_advisor"),
];

const travelAdvisor = createReactAgent({
  llm: model,
  tools: travelAdvisorTools,
  prompt: [
    "You are a general travel expert that can recommend travel destinations (e.g. countries, cities, etc). ",
    "If you need hotel recommendations, ask 'hotel_advisor' for help. ",
    "You MUST include human-readable response before transferring to another agent."
  ].join("")
});

const callTravelAdvisor = async (
  state: z.infer<typeof MultiAgentState>
): Promise<Command> => {
  const response = await travelAdvisor.invoke(state);
  const update = { ...response, lastActiveAgent: "travel_advisor" };
  return new Command({ update, goto: "human" });
};

// Define hotel advisor tools
const getHotelRecommendations = tool(
  async () => {
    // Placeholder implementation
    return "I recommend the Ritz-Carlton for luxury stays or boutique hotels for unique experiences.";
  },
  {
    name: "get_hotel_recommendations",
    description: "Get hotel recommendations for destinations",
    schema: z.object({}),
  }
);

const hotelAdvisorTools = [
  getHotelRecommendations,
  makeHandoffTool("travel_advisor"),
];

const hotelAdvisor = createReactAgent({
  llm: model,
  tools: hotelAdvisorTools,
  prompt: [
    "You are a hotel expert that can provide hotel recommendations for a given destination. ",
    "If you need help picking travel destinations, ask 'travel_advisor' for help.",
    "You MUST include human-readable response before transferring to another agent."
  ].join("")
});

const callHotelAdvisor = async (
  state: z.infer<typeof MultiAgentState>
): Promise<Command> => {
  const response = await hotelAdvisor.invoke(state);
  const update = { ...response, lastActiveAgent: "hotel_advisor" };
  return new Command({ update, goto: "human" });
};

const humanNode = async (
  state: z.infer<typeof MultiAgentState>
): Promise<Command> => {
  const userInput: string = interrupt("Ready for user input.");
  const activeAgent = state.lastActiveAgent || "travel_advisor";

  return new Command({
    update: {
      messages: [
        {
          role: "human",
          content: userInput,
        }
      ]
    },
    goto: activeAgent,
  });
};

const builder = new StateGraph(MultiAgentState)
  .addNode("travel_advisor", callTravelAdvisor)
  .addNode("hotel_advisor", callHotelAdvisor)
  .addNode("human", humanNode)
  .addEdge(START, "travel_advisor");

const checkpointer = new MemorySaver();
const graph = builder.compile({ checkpointer });

Let’s test a multi turn conversation with this application.

import { v4 as uuidv4 } from "uuid";
import { Command } from "@langchain/langgraph";

const threadConfig = { configurable: { thread_id: uuidv4() } };

const inputs = [
  // 1st round of conversation
  {
    messages: [
      { role: "user", content: "i wanna go somewhere warm in the caribbean" }
    ]
  },
  // Since we're using `interrupt`, we'll need to resume using the Command primitive.
  // 2nd round of conversation
  new Command({
    resume: "could you recommend a nice hotel in one of the areas and tell me which area it is."
  }),
  // 3rd round of conversation
  new Command({
    resume: "i like the first one. could you recommend something to do near the hotel?"
  }),
];

for (const [idx, userInput] of inputs.entries()) {
  console.log();
  console.log(`--- Conversation Turn ${idx + 1} ---`);
  console.log();
  console.log(`User: ${JSON.stringify(userInput)}`);
  console.log();
  
  for await (const update of await graph.stream(
    userInput,
    { ...threadConfig, streamMode: "updates" }
  )) {
    for (const [nodeId, value] of Object.entries(update)) {
      if (value?.messages?.length) {
        const lastMessage = value.messages.at(-1);
        if (lastMessage?.getType?.() === "ai") {
          console.log(`${nodeId}: ${lastMessage.content}`);
        }
      }
    }
  }
}

--- Conversation Turn 1 ---

User: {"messages":[{"role":"user","content":"i wanna go somewhere warm in the caribbean"}]}

travel_advisor: Based on the recommendations, Aruba would be an excellent choice for your Caribbean getaway! Aruba is known as "One Happy Island" and offers:
- Year-round warm weather with consistent temperatures around 82°F (28°C)
- Beautiful white sand beaches like Eagle Beach and Palm Beach
- Clear turquoise waters perfect for swimming and snorkeling
- Minimal rainfall and location outside the hurricane belt
- A blend of Caribbean and Dutch culture
- Great dining options and nightlife
- Various water sports and activities

Would you like me to get some specific hotel recommendations in Aruba for your stay? I can transfer you to our hotel advisor who can help with accommodations.

--- Conversation Turn 2 ---

User: Command { resume: 'could you recommend a nice hotel in one of the areas and tell me which area it is.' }

hotel_advisor: Based on the recommendations, I can suggest two excellent options:

1. The Ritz-Carlton, Aruba - Located in Palm Beach
- This luxury resort is situated in the vibrant Palm Beach area
- Known for its exceptional service and amenities
- Perfect if you want to be close to dining, shopping, and entertainment
- Features multiple restaurants, a casino, and a world-class spa
- Located on a pristine stretch of Palm Beach

2. Bucuti & Tara Beach Resort - Located in Eagle Beach
- An adults-only boutique resort on Eagle Beach
- Known for being more intimate and peaceful
- Award-winning for its sustainability practices
- Perfect for a romantic getaway or peaceful vacation
- Located on one of the most beautiful beaches in the Caribbean

Would you like more specific information about either of these properties or their locations?

--- Conversation Turn 3 ---

User: Command { resume: 'i like the first one. could you recommend something to do near the hotel?' }

travel_advisor: Near the Ritz-Carlton in Palm Beach, here are some highly recommended activities:

1. Visit the Palm Beach Plaza Mall - Just a short walk from the hotel, featuring shopping, dining, and entertainment
2. Try your luck at the Stellaris Casino - It's right in the Ritz-Carlton
3. Take a sunset sailing cruise - Many depart from the nearby pier
4. Visit the California Lighthouse - A scenic landmark just north of Palm Beach
5. Enjoy water sports at Palm Beach:
   - Jet skiing
   - Parasailing
   - Snorkeling
   - Stand-up paddleboarding

Would you like more specific information about any of these activities or would you like to know about other options in the area?

Prebuilt implementations

LangGraph comes with prebuilt implementations of two of the most popular multi-agent architectures:

supervisor — individual agents are coordinated by a central supervisor agent. The supervisor controls all communication flow and task delegation, making decisions about which agent to invoke based on the current context and task requirements. You can use langgraph-supervisor library to create a supervisor multi-agent systems.
swarm — agents dynamically hand off control to one another based on their specializations. The system remembers which agent was last active, ensuring that on subsequent interactions, the conversation resumes with that agent. You can use langgraph-swarm library to create a swarm multi-agent systems.

Basic configuration

Low-level configuration

Components

Build custom multi-agent systems

Handoffs

Create handoffs

Control agent inputs

Build a multi-agent system

Multi-turn conversation

Prebuilt implementations

Basic configuration

Low-level configuration

Components

​Handoffs

​Create handoffs

​Control agent inputs

​Build a multi-agent system

​Multi-turn conversation

​Prebuilt implementations

Handoffs

Create handoffs

Control agent inputs

Build a multi-agent system

Multi-turn conversation

Prebuilt implementations