多智能体协作实战：AutoGen / CrewAI / LangGraph 选型指南 Multi-Agent Collaboration: AutoGen vs CrewAI vs LangGraph

什么是多智能体协作？ What Is Multi-Agent Collaboration?

单个 AI Agent 能做很多事情，但它有一个根本限制：它只能扮演一个角色。如果你需要一篇文章先由研究员搜集资料，再由写作者撰写初稿，再由编辑审核修改——单 Agent 做不到这种多角色的流水线协作。多智能体协作（Multi-Agent）就是让多个 AI Agent 各自扮演不同角色，按照一定的流程协作完成复杂任务。 A single AI Agent can do many things, but it has a fundamental limitation: it can only play one role. If you need an article first researched, then drafted by a writer, then reviewed and revised by an editor — a single Agent can't handle this multi-role pipeline. Multi-Agent collaboration means having multiple AI Agents each play different roles, collaborating through defined workflows to complete complex tasks.

多智能体不是"同时用多个 AI 聊天"，而是让不同的 Agent 有明确的分工、明确的输入输出、明确的协作流程。就像一个公司里，产品经理、设计师、工程师各司其职，通过流程协作完成一个项目。2025 年，三个主流框架——AutoGen、CrewAI、LangGraph——让这种协作变得可编程、可复用。 Multi-Agent isn't "using multiple AIs to chat at the same time" — it's giving different Agents clear roles, clear inputs/outputs, and clear collaboration workflows. Like a company where product managers, designers, and engineers each have their responsibilities, collaborating through processes to complete a project. In 2025, three mainstream frameworks — AutoGen, CrewAI, and LangGraph — make this collaboration programmable and reusable.

三大框架对比 Framework Comparison

LangGraph（LangChain）——单独说是因为它的定位不同。LangGraph 不是"多 Agent 框架"，而是基于图的状态机框架。你可以用它构建任意复杂的 Agent 工作流，包括多 Agent 协作。学习曲线：最高，但灵活度也最高。适合需要精细控制每一步流程、需要状态持久化、需要人机协作循环的复杂场景。如果你需要构建生产级的 Agent 系统，LangGraph 是最可靠的选择。 LangGraph (LangChain) — discussed separately because its positioning differs. LangGraph isn't a "multi-Agent framework" but a graph-based state machine framework. You can build arbitrarily complex Agent workflows with it, including multi-Agent collaboration. Learning curve: highest, but so is flexibility. Best for complex scenarios needing fine-grained flow control, state persistence, and human-AI collaboration loops. If you're building production-grade Agent systems, LangGraph is the most reliable choice.

实战：用 CrewAI 做一个内容创作流水线 Hands-On: Build a Content Pipeline with CrewAI

选 CrewAI 做示例，因为它的代码最简洁直观。我们要做一个"技术博客写作流水线"：研究员搜集资料、写作者写初稿、编辑审核修改。以下是完整代码： Choosing CrewAI for the example because its code is the most concise and intuitive. We'll build a "tech blog writing pipeline": a researcher gathers info, a writer drafts, an editor reviews. Here's the complete code:

from crewai import Agent, Task, Crew, Process
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(model="gpt-4o", temperature=0.7)

# 定义三个 Agent
researcher = Agent(
    role="技术研究员",
    goal="搜集关于指定主题的最新技术资料",
    backstory="你是一名资深技术研究员，擅长快速搜集和分析技术信息",
    llm=llm,
    verbose=True
)

writer = Agent(
    role="技术写作者",
    goal="基于研究员提供的资料，撰写高质量的技术博客文章",
    backstory="你是一名经验丰富的技术博主，文章风格简洁清晰、有深度",
    llm=llm,
    verbose=True
)

editor = Agent(
    role="技术编辑",
    goal="审核并优化文章质量，确保准确性和可读性",
    backstory="你是一名严谨的技术编辑，注重事实准确和表达清晰",
    llm=llm,
    verbose=True
)

# 定义三个任务
research_task = Task(
    description="搜集关于 '{topic}' 的最新技术资料，整理成要点清单",
    expected_output="一份包含 5-8 个关键要点的资料清单",
    agent=researcher
)

writing_task = Task(
    description="基于研究员提供的资料清单，撰写一篇 1500 字的技术博客",
    expected_output="一篇结构完整、有深度的技术博客文章",
    agent=writer
)

editing_task = Task(
    description="审核文章的技术准确性、逻辑结构和表达质量，给出修改建议并输出最终版本",
    expected_output="经过审核和优化后的最终版文章",
    agent=editor
)

# 组装 Crew 并执行
crew = Crew(
    agents=[researcher, writer, editor],
    tasks=[research_task, writing_task, editing_task],
    process=Process.sequential,  # 顺序执行
    verbose=True
)

result = crew.kickoff(inputs={"topic": "2025 年多智能体框架对比"})
print(result)

这段代码定义了三个 Agent（研究员、写作者、编辑）和三个 Task（搜集资料、写文章、审校），然后按顺序执行。CrewAI 会自动把前一个 Task 的输出作为后一个 Task 的输入。整个多 Agent 协作系统，核心代码不超过 50 行。 This code defines three Agents (researcher, writer, editor) and three Tasks (gather info, write article, review), then executes them sequentially. CrewAI automatically passes the previous Task's output as the next Task's input. The entire multi-Agent collaboration system, core code under 50 lines.

从零搭建的 5 个步骤 5 Steps to Build from Scratch