A Python library for creating swarm-style multi-agent systems using LangGraph. A swarm is a type of multi-agent architecture where 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.
This library is built on top of LangGraph, a powerful framework for building agent applications, and comes with out-of-box support for streaming, short-term and long-term memory and human-in-the-loop
pip install langgraph-swarmpip install langgraph-swarm langchain-openai
export OPENAI_API_KEY=<your_api_key>from langchain_openai import ChatOpenAI
from langgraph.checkpoint.memory import InMemorySaver
from langchain.agents import create_agent
from langgraph_swarm import create_handoff_tool, create_swarm
model = ChatOpenAI(model="gpt-4o")
def add(a: int, b: int) -> int:
"""Add two numbers"""
return a + b
alice = create_agent(
model,
tools=[
add,
create_handoff_tool(
agent_name="Bob",
description="Transfer to Bob",
),
],
system_prompt="You are Alice, an addition expert.",
name="Alice",
)
bob = create_agent(
model,
tools=[
create_handoff_tool(
agent_name="Alice",
description="Transfer to Alice, she can help with math",
),
],
system_prompt="You are Bob, you speak like a pirate.",
name="Bob",
)
checkpointer = InMemorySaver()
workflow = create_swarm(
[alice, bob],
default_active_agent="Alice"
)
app = workflow.compile(checkpointer=checkpointer)
config = {"configurable": {"thread_id": "1"}}
turn_1 = app.invoke(
{"messages": [{"role": "user", "content": "i'd like to speak to Bob"}]},
config,
)
print(turn_1)
turn_2 = app.invoke(
{"messages": [{"role": "user", "content": "what's 5 + 7?"}]},
config,
)
print(turn_2)[!TIP] For developing, debugging, and deploying AI agents and LLM applications, see LangSmith.
You can add short-term and long-term memory to your swarm multi-agent system. Since create_swarm() returns an instance of StateGraph that needs to be compiled before use, you can directly pass a checkpointer or a store instance to the .compile() method:
from langgraph.checkpoint.memory import InMemorySaver
from langgraph.store.memory import InMemoryStore
# short-term memory
checkpointer = InMemorySaver()
# long-term memory
store = InMemoryStore()
model = ...
alice = ...
bob = ...
workflow = create_swarm(
[alice, bob],
default_active_agent="Alice"
)
# Compile with checkpointer/store
app = workflow.compile(
checkpointer=checkpointer,
store=store
)[!IMPORTANT] Adding short-term memory is crucial for maintaining conversation state across multiple interactions. Without it, the swarm would "forget" which agent was last active and lose the conversation history. Make sure to always compile the swarm with a checkpointer if you plan to use it in multi-turn conversations; e.g.,
workflow.compile(checkpointer=checkpointer).
You can customize multi-agent swarm by changing either the handoff tools implementation or the agent implementation.
By default, the agents in the swarm are assumed to use handoff tools created with the prebuilt create_handoff_tool. You can also create your own, custom handoff tools. Here are some ideas on how you can modify the default implementation:
create_handoff_tool passes full message history (all of the messages generated in the swarm up to this point), as well as a tool message indicating successful handoff.Here is an example of what a custom handoff tool might look like:
from typing import Annotated
from langchain.tools import tool, BaseTool, InjectedToolCallId
from langchain.messages import ToolMessage
from langgraph.types import Command
from langgraph.prebuilt import InjectedState
def create_custom_handoff_tool(*, agent_name: str, name: str | None, description: str | None) -> BaseTool:
@tool(name, description=description)
def handoff_to_agent(
# you can add additional tool call arguments for the LLM to populate
# for example, you can ask the LLM to populate a task description for the next agent
task_description: Annotated[str, "Detailed description of what the next agent should do, including all of the relevant context."],
# you can inject the state of the agent that is calling the tool
state: Annotated[dict, InjectedState],
tool_call_id: Annotated[str, InjectedToolCallId],
):
tool_message = ToolMessage(
content=f"Successfully transferred to {agent_name}",
name=name,
tool_call_id=tool_call_id,
)
# you can use a different messages state key here, if your agent uses a different schema
# e.g., "alice_messages" instead of "messages"
messages = state["messages"]
return Command(
goto=agent_name,
graph=Command.PARENT,
# NOTE: this is a state update that will be applied to the swarm multi-agent graph (i.e., the PARENT graph)
update={
"messages": messages + [tool_message],
"active_agent": agent_name,
# optionally pass the task description to the next agent
"task_description": task_description,
},
)
return handoff_to_agent[!IMPORTANT] If you are implementing custom handoff tools that return
Command, you need to ensure that:
(1) your agent has a tool-calling node that can handle tools returningCommand(like LangGraph's prebuiltToolNode)
(2) both the swarm graph and the next agent graph have the state schema containing the keys you want to update inCommand.update
By default, individual agents are expected to communicate over a single messages key that is shared by all agents and the overall multi-agent swarm graph. This means that messages from all of the agents will be combined into a single, shared list of messages. This might not be desirable if you don't want to expose an agent's internal history of messages. To change this, you can customize the agent by taking the following steps:
alice_messagesfrom typing_extensions import TypedDict, Annotated
from langchain.messages import AnyMessage
from langgraph.graph import StateGraph, add_messages
from langgraph_swarm import SwarmState
class AliceState(TypedDict):
alice_messages: Annotated[list[AnyMessage], add_messages]
# see this guide to learn how you can implement a custom tool-calling agent
# https://langchain-ai.github.io/langgraph/how-tos/react-agent-from-scratch/
alice = (
StateGraph(AliceState)
.add_node("model", ...)
.add_node("tools", ...)
.add_edge(...)
...
.compile()
)
# wrapper calling the agent
def call_alice(state: SwarmState):
# you can put any input transformation from parent state -> agent state
# for example, you can invoke "alice" with "task_description" populated by the LLM
response = alice.invoke({"alice_messages": state["messages"]})
# you can put any output transformation from agent state -> parent state
return {"messages": response["alice_messages"]}
def call_bob(state: SwarmState):
...Then, you can create the swarm manually in the following way:
from langgraph_swarm import add_active_agent_router
workflow = (
StateGraph(SwarmState)
.add_node("Alice", call_alice, destinations=("Bob",))
.add_node("Bob", call_bob, destinations=("Alice",))
)
# this is the router that enables us to keep track of the last active agent
workflow = add_active_agent_router(
builder=workflow,
route_to=["Alice", "Bob"],
default_active_agent="Alice",
)
# compile the workflow
app = workflow.compile()Create a tool that can handoff control to the requested agent.
Get a list of destinations from agent's handoff tools.
Add a router to the currently active agent to the StateGraph.
Create a multi-agent swarm.