LangGraph 智能体状态管理与决策

简介

本次演示围绕 Bright Data Web MCP 与 LangGraph 的集成实操展开，完整展示了从获取大模型 API Key、创建大模型会话，到获取 Bright Data API Key、通过 MultiServerMCPClient 连接 Web MCP 服务器，并在 Bright Data 后台进一步启用浏览器自动化工具、扩展智能体可调用能力的全流程；同时结合 LangGraph 的 StateGraph，搭建了包含大模型节点、工具调用节点和路由规则节点的循环式 AI 研究智能体。演示过程中，通过'打开网页并持续滚动，直到提取 30 条语录的作者、内容与标签'这一实际任务，直观呈现了智能体基于实时网页数据进行搜索、抓取、交互和推理的完整效果。实测结果表明，LangGraph 提供了清晰可控的智能体状态管理与决策机制，而 Bright Data Web MCP 则补足了真实网页访问与动态页面交互能力，使 AI Agent 无需将复杂抓取逻辑硬编码进提示词或业务代码中，也能更稳定地完成生产级研究任务。

from __future__ import annotations
import argparse
import asyncio
import json
import os
import sys
from typing import Any, Literal
from urllib.parse import urlencode
from dotenv import load_dotenv
from langchain_core.messages import HumanMessage, SystemMessage, ToolMessage
from langchain_openai import ChatOpenAI
from langchain_mcp_adapters.client import MultiServerMCPClient
from langgraph.graph import END, START, MessagesState, StateGraph

SYSTEM_PROMPT = """
You are a web research assistant. Task:
- Research the user's topic using Google search results and a few sources.
- Return 6–10 simple bullet points.
- Add a short "Sources:" list with only the URLs you used.
How to use tools:
- First call the search tool to get Google results.
- Select 3–5 reputable results and scrape them.
- If scraping fails, try a different result.
Constraints:
- Use at most 5 sources.
- Prefer official docs or primary sources.
- Keep it quick: no deep crawling.
"""

def make_llm_call_node(llm_with_tools):
    async def llm_call():
        messages = [SystemMessage(content=SYSTEM_PROMPT)] + state[]
        ai_message =  llm_with_tools.ainvoke(messages)
         {: [ai_message]}
     llm_call

 ():
      ():
        last_ai_msg = state[][-]
        tool_results = []
         tool_call  last_ai_msg.tool_calls:
            tool = tools_by_name.get(tool_call[])
              tool:
                tool_results.append(
                    ToolMessage(
                        content=,
                        tool_call_id=tool_call[],
                    )
                )
                
            observation = (
                 tool.ainvoke(tool_call[])
                 (tool, )
                 tool.invoke(tool_call[])
            )
            tool_results.append(
                ToolMessage(
                    content=(observation),
                    tool_call_id=tool_call[],
                )
            )
         {: tool_results}
     tool_node

 () -> [, END]:
    last_message = state[][-]
     (last_message, , ):
         
     END

  ():
    load_dotenv()
    bd_token = os.getenv()
      bd_token:
         ValueError()

    client = MultiServerMCPClient({
        : {
            : ,
            : ,
        }
    })

    tools =  client.get_tools()
    tools_by_name = {tool.name: tool  tool  tools}
    ()

    openai_api_key = os.getenv()
    llm = ChatOpenAI(
        model=,
        temperature=,
        api_key=openai_api_key,
        base_url=,
    )
    llm_with_tools = llm.bind_tools(tools)

    graph = StateGraph(MessagesState)
    graph.add_node(, make_llm_call_node(llm_with_tools))
    graph.add_node(, make_tool_node(tools_by_name))

    graph.add_edge(START, )
    graph.add_conditional_edges(, should_continue, [, END])
    graph.add_edge(, )
    agent = graph.()

    topic = 
    result =  agent.ainvoke(
        {: [HumanMessage(content=)]},
        config={: }
    )
    (result[][-].content)

 __name__ == :
    asyncio.run(main())