构建多智能体系统 - LangChain 框架 --知识铺

构建多智能体系统¶

如果单个智能体需要在多个领域进行专业化或管理许多工具，它可能会面临困难。为了解决这个问题，您可以将智能体分解为更小、独立的智能体，并将它们组合成一个多智能体系统。

在多智能体系统中，智能体之间需要相互通信。它们通过交接实现这一点——交接是一种原语，描述了将控制权移交给哪个智能体以及要发送给该智能体的数据负载。

本指南涵盖以下内容

实现智能体之间的交接
使用交接和预构建的智能体来构建自定义多智能体系统

要开始构建多智能体系统，请查看 LangGraph 预构建的两种最流行的多智能体架构实现——监督者和群组。

交接¶

为了在多智能体系统中设置智能体之间的通信，您可以使用交接（handoffs）——这是一种一个智能体将控制权“交接”给另一个智能体的模式。交接允许您指定

目标：要导航到的目标智能体（例如，要前往的 LangGraph 节点的名称）
数据负载：要传递给该智能体的信息（例如，状态更新）

创建交接¶

要实现交接，您可以从智能体节点或工具中返回 Command 对象

<span id="__span-0-1">from typing import Annotated
<span id="__span-0-2">from langchain_core.tools import tool, InjectedToolCallId
<span id="__span-0-3">from langgraph.prebuilt import create_react_agent, InjectedState
<span id="__span-0-4">from langgraph.graph import StateGraph, START, MessagesState
<span id="__span-0-5">from langgraph.types import Command
<span id="__span-0-6">
<span id="__span-0-7">def create_handoff_tool(*, agent_name: str, description: str | None = None):
<span id="__span-0-8">    name = f"transfer_to_{agent_name}"
<span id="__span-0-9">    description = description or f"Transfer to {agent_name}"
<span id="__span-0-10">
<span id="__span-0-11">    @tool(name, description=description)
<span id="__span-0-12">    def handoff_tool(
<span id="__span-0-13">        state: Annotated[MessagesState, InjectedState], 
<span id="__span-0-14">        tool_call_id: Annotated[str, InjectedToolCallId],
<span id="__span-0-15">    ) -&gt; Command:
<span id="__span-0-16">        tool_message = {
<span id="__span-0-17">            "role": "tool",
<span id="__span-0-18">            "content": f"Successfully transferred to {agent_name}",
<span id="__span-0-19">            "name": name,
<span id="__span-0-20">            "tool_call_id": tool_call_id,
<span id="__span-0-21">        }
<span id="__span-0-22">        return Command(  
<span id="__span-0-23">            goto=agent_name,  
<span id="__span-0-24">            update={"messages": state["messages"] + [tool_message]},  
<span id="__span-0-25">            graph=Command.PARENT,  
<span id="__span-0-26">        )
<span id="__span-0-27">    return handoff_tool

提示

如果您想使用返回 Command 的工具，您可以使用预构建的 [create_react_agent](https://langgraph.com.cn/reference/agents.1.html#langgraph.prebuilt.chat_agent_executor.create_react_agent “ create_react_agent”) / [ToolNode](https://langgraph.com.cn/reference/agents.1.html#langgraph.prebuilt.tool_node.ToolNode “ ToolNode”) 组件，或者实现自己的工具执行节点，该节点收集工具返回的 Command 对象并返回它们的列表，例如。

<span id="__span-1-1">def call_tools(state):
<span id="__span-1-2">    ...
<span id="__span-1-3">    commands = [tools_by_name[tool_call["name"]].invoke(tool_call) for tool_call in tool_calls]
<span id="__span-1-4">    return commands

重要

此交接实现假定

多智能体系统中的每个智能体都将整体消息历史（跨所有智能体）作为其输入。如果您想更精细地控制智能体输入，请参阅此部分

每个智能体将其内部消息历史输出到多智能体系统的整体消息历史中。如果您想更精细地控制**智能体输出的添加方式**，请将智能体包装在一个单独的节点函数中

<span id="__span-2-1">def call_hotel_assistant(state):
<span id="__span-2-2">    # return agent's final response,
<span id="__span-2-3">    # excluding inner monologue
<span id="__span-2-4">    response = hotel_assistant.invoke(state)
<span id="__span-2-5">    return {"messages": response["messages"][-1]}

控制智能体输入¶

您可以使用 [Send()](https://langgraph.com.cn/reference/types/index.html#langgraph.types.Send “ Send”) 原语在交接期间直接向工作智能体发送数据。例如，您可以要求调用智能体为下一个智能体填充任务描述

<code tabindex="0"><span id="__span-3-1">from typing import Annotated
<span id="__span-3-2">from langchain_core.tools import tool, InjectedToolCallId
<span id="__span-3-3">from langgraph.prebuilt import InjectedState
<span id="__span-3-4">from langgraph.graph import StateGraph, START, MessagesState
<span id="__span-3-5">from langgraph.types import Command, Send
<span id="__span-3-6">
<span id="__span-3-7">def create_task_description_handoff_tool(
<span id="__span-3-8">    *, agent_name: str, description: str | None = None
<span id="__span-3-9">):
<span id="__span-3-10">    name = f"transfer_to_{agent_name}"
<span id="__span-3-11">    description = description or f"Ask {agent_name} for help."
<span id="__span-3-12">
<span id="__span-3-13">    @tool(name, description=description)
<span id="__span-3-14">    def handoff_tool(
<span id="__span-3-15">        # this is populated by the calling agent
<span id="__span-3-16">        task_description: Annotated[
<span id="__span-3-17">            str,
<span id="__span-3-18">            "Description of what the next agent should do, including all of the relevant context.",
<span id="__span-3-19">        ],
<span id="__span-3-20">        # these parameters are ignored by the LLM
<span id="__span-3-21">        state: Annotated[MessagesState, InjectedState],
<span id="__span-3-22">    ) -&gt; Command:
<span id="__span-3-23">        task_description_message = {"role": "user", "content": task_description}
<span id="__span-3-24">        agent_input = {**state, "messages": [task_description_message]}
<span id="__span-3-25">        return Command(
<span id="__span-3-26">            goto=[Send(agent_name, agent_input)],
<span id="__span-3-27">            graph=Command.PARENT,
<span id="__span-3-28">        )
<span id="__span-3-29">
<span id="__span-3-30">    return handoff_tool

有关在交接中使用 [Send()](https://langgraph.com.cn/reference/types/index.html#langgraph.types.Send “ Send”) 的完整示例，请参阅多智能体监督者教程。