Python 构建AI多智能体系统：让三个 AI 协作完成复杂任务

单个 AI 能做的事有限，三个 AI 分工协作能做的事远超你想象。本文用纯 Python 从零实现一个多智能体系统，完整代码可直接运行。

• • 一、什么是多智能体系统（Multi-Agent System）？
  • 二、为什么不用现成框架？
  • 三、系统架构设计
  • 四、代码实现
  • • 4.1 基础类：定义 Agent
    • 4.2 定义三个专业 Agent
    • 4.3 协调者：任务调度中心
    • 4.4 主程序：启动你的 AI 团队
  • 五、执行流程详解
  • 六、扩展：加入工具能力
  • 七、性能对比
  • 八、注意事项
  • 总结

一、什么是多智能体系统（Multi-Agent System）？

简单说：多个 AI Agent 各司其职，协作完成一个复杂任务。

类比现实团队：

• 产品经理负责分析需求
• 程序员负责写代码
• 测试工程师负责检查质量

每个 Agent 有自己的职责、工具和思考方式，它们之间通过消息传递来协作。

#mermaid-svg-n1hhyDoBVcJeh7iS{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-n1hhyDoBVcJeh7iS .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-n1hhyDoBVcJeh7iS .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-n1hhyDoBVcJeh7iS .error-icon{fill:#552222;}#mermaid-svg-n1hhyDoBVcJeh7iS .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-n1hhyDoBVcJeh7iS .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-n1hhyDoBVcJeh7iS .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-n1hhyDoBVcJeh7iS .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-n1hhyDoBVcJeh7iS .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-n1hhyDoBVcJeh7iS .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-n1hhyDoBVcJeh7iS .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-n1hhyDoBVcJeh7iS .marker{fill:#333333;stroke:#333333;}#mermaid-svg-n1hhyDoBVcJeh7iS .marker.cross{stroke:#333333;}#mermaid-svg-n1hhyDoBVcJeh7iS svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-n1hhyDoBVcJeh7iS p{margin:0;}#mermaid-svg-n1hhyDoBVcJeh7iS .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-n1hhyDoBVcJeh7iS .cluster-label text{fill:#333;}#mermaid-svg-n1hhyDoBVcJeh7iS .cluster-label span{color:#333;}#mermaid-svg-n1hhyDoBVcJeh7iS .cluster-label span p{background-color:transparent;}#mermaid-svg-n1hhyDoBVcJeh7iS .label text,#mermaid-svg-n1hhyDoBVcJeh7iS span{fill:#333;color:#333;}#mermaid-svg-n1hhyDoBVcJeh7iS .node rect,#mermaid-svg-n1hhyDoBVcJeh7iS .node circle,#mermaid-svg-n1hhyDoBVcJeh7iS .node ellipse,#mermaid-svg-n1hhyDoBVcJeh7iS .node polygon,#mermaid-svg-n1hhyDoBVcJeh7iS .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-n1hhyDoBVcJeh7iS .rough-node .label text,#mermaid-svg-n1hhyDoBVcJeh7iS .node .label text,#mermaid-svg-n1hhyDoBVcJeh7iS .image-shape .label,#mermaid-svg-n1hhyDoBVcJeh7iS .icon-shape .label{text-anchor:middle;}#mermaid-svg-n1hhyDoBVcJeh7iS .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-n1hhyDoBVcJeh7iS .rough-node .label,#mermaid-svg-n1hhyDoBVcJeh7iS .node .label,#mermaid-svg-n1hhyDoBVcJeh7iS .image-shape .label,#mermaid-svg-n1hhyDoBVcJeh7iS .icon-shape .label{text-align:center;}#mermaid-svg-n1hhyDoBVcJeh7iS .node.clickable{cursor:pointer;}#mermaid-svg-n1hhyDoBVcJeh7iS .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-n1hhyDoBVcJeh7iS .arrowheadPath{fill:#333333;}#mermaid-svg-n1hhyDoBVcJeh7iS .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-n1hhyDoBVcJeh7iS .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-n1hhyDoBVcJeh7iS .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-n1hhyDoBVcJeh7iS .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-n1hhyDoBVcJeh7iS .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-n1hhyDoBVcJeh7iS .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-n1hhyDoBVcJeh7iS .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-n1hhyDoBVcJeh7iS .cluster text{fill:#333;}#mermaid-svg-n1hhyDoBVcJeh7iS .cluster span{color:#333;}#mermaid-svg-n1hhyDoBVcJeh7iS div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-n1hhyDoBVcJeh7iS .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-n1hhyDoBVcJeh7iS rect.text{fill:none;stroke-width:0;}#mermaid-svg-n1hhyDoBVcJeh7iS .icon-shape,#mermaid-svg-n1hhyDoBVcJeh7iS .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-n1hhyDoBVcJeh7iS .icon-shape p,#mermaid-svg-n1hhyDoBVcJeh7iS .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-n1hhyDoBVcJeh7iS .icon-shape .label rect,#mermaid-svg-n1hhyDoBVcJeh7iS .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-n1hhyDoBVcJeh7iS .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-n1hhyDoBVcJeh7iS .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-n1hhyDoBVcJeh7iS :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

调研结果

代码输出

反馈意见

需要修改

📋 最终交付

🎯 协调者 Agent任务拆分与调度

🔍 研究 Agent信息搜集与分析

💻 编码 Agent代码生成

🧪 审查 Agent质量检查

二、为什么不用现成框架？

LangGraph、CrewAI、AutoGen 都很好，但先理解原理再选框架，才是正确的学习路径。

#mermaid-svg-31fHGVM4EzwqBz4k{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-31fHGVM4EzwqBz4k .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-31fHGVM4EzwqBz4k .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-31fHGVM4EzwqBz4k .error-icon{fill:#552222;}#mermaid-svg-31fHGVM4EzwqBz4k .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-31fHGVM4EzwqBz4k .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-31fHGVM4EzwqBz4k .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-31fHGVM4EzwqBz4k .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-31fHGVM4EzwqBz4k .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-31fHGVM4EzwqBz4k .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-31fHGVM4EzwqBz4k .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-31fHGVM4EzwqBz4k .marker{fill:#333333;stroke:#333333;}#mermaid-svg-31fHGVM4EzwqBz4k .marker.cross{stroke:#333333;}#mermaid-svg-31fHGVM4EzwqBz4k svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-31fHGVM4EzwqBz4k p{margin:0;}#mermaid-svg-31fHGVM4EzwqBz4k .pieCircle{stroke:#000000;stroke-width:2px;opacity:0.7;}#mermaid-svg-31fHGVM4EzwqBz4k .pieOuterCircle{stroke:#000000;stroke-width:1px;fill:none;}#mermaid-svg-31fHGVM4EzwqBz4k .pieTitleText{text-anchor:middle;font-size:25px;fill:#000000;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}#mermaid-svg-31fHGVM4EzwqBz4k .slice{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;fill:#000000;font-size:17px;}#mermaid-svg-31fHGVM4EzwqBz4k .legend text{fill:#000000;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:17px;}#mermaid-svg-31fHGVM4EzwqBz4k :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

32%

28%

22%

13%

5%

多智能体开发者的技术选型（2026）

自研轻量框架

LangGraph

CrewAI

AutoGen

其他

本文先用纯 Python 实现核心逻辑，再对比框架选型。

三、系统架构设计

我们要构建的系统：AI 研发小队

用户输入需求 → 协调者拆分任务 → 三位 Agent 各司其职 → 汇总交付

四、代码实现

4.1 基础类：定义 Agent

# agent.py
import json
import ollama
from dataclasses import dataclass, field
from typing import Optional

@dataclass
class Message:
role: str # "user" | "agent" | "system"
sender: str # 发送者名称
content: str

@dataclass
class Agent:
name: str
role: str
system_prompt: str
model: str = "qwen3:8b"
history: list = field(default_factory=list)

def think(self, messages: list[Message]) –> str:
"""思考并回复"""
# 构建完整的消息列表
full_messages = [
{"role": "system", "content": self.system_prompt}
]
for msg in messages:
full_messages.append({
"role": "user" if msg.sender != self.name else "assistant",
"content": f"[来自 {msg.sender}] {msg.content}"
})

# 调用模型
response = ollama.chat(
model=self.model,
messages=full_messages
)
return response["message"]["content"]

4.2 定义三个专业 Agent

# agents.py
from agent import Agent

# 研究员 Agent：负责技术调研
researcher = Agent(
name="Researcher",
role="技术研究员",
system_prompt="""你是一个技术研究员。你的任务是：
1. 分析用户需求中涉及的技术点
2. 给出最佳实践和技术方案建议
3. 列出关键注意事项和潜在风险

输出格式：
## 技术分析
（分析内容）

## 推荐方案
（方案详情）

## 注意事项
（风险和注意点）"""
)

# 程序员 Agent：负责写代码
coder = Agent(
name="Coder",
role="高级程序员",
system_prompt="""你是一个高级 Python 程序员。你的任务是：
1. 根据需求和调研结果编写高质量的 Python 代码
2. 代码必须包含完整的类型注解和 docstring
3. 包含错误处理和日志记录
4. 输出可直接运行的完整代码

输出格式：
## 实现思路
（简述思路）

## 完整代码
```python
（代码）
```"""
)

# 审查员 Agent：负责代码审查
reviewer = Agent(
name="Reviewer",
role="代码审查专家",
system_prompt="""你是一个严格的代码审查专家。你的任务是：
1. 审查代码的正确性、安全性和性能
2. 检查是否遵循最佳实践
3. 给出修改建议

评分标准（满分 10 分）：
– 正确性：代码逻辑是否正确
– 安全性：是否存在安全漏洞
– 可读性：命名、注释、代码结构
– 性能：是否有明显的性能问题

输出格式：
## 评分：X/10
## 优点
（列出优点）
## 问题
（列出问题）
## 修改建议
（具体建议）"""
)

4.3 协调者：任务调度中心

# coordinator.py
from agent import Agent, Message

class Coordinator:
"""协调者：负责调度多个 Agent 协作"""

def __init__(self, researcher, coder, reviewer):
self.researcher = researcher
self.coder = coder
self.reviewer = reviewer
self.log = []

def _log(self, phase: str, content: str):
"""记录执行日志"""
self.log.append({"phase": phase, "content": content})
print(f"\\n{'='*60}")
print(f"📋 [{phase}]")
print(f"{'='*60}")
print(content[:500] + "…" if len(content) > 500 else content)

def run(self, task: str, max_revisions: int = 2) –> dict:
"""
执行完整的多智能体协作流程

Args:
task: 用户需求描述
max_revisions: 最大修改轮数

Returns:
包含所有阶段输出的字典
"""
results = {"task": task, "phases": []}

# ========== 第一阶段：需求分析 + 技术调研 ==========
print("\\n🔍 第一阶段：技术调研")
research_msg = Message(
role="user", sender="Coordinator",
content=f"请分析以下需求的技术方案：\\n{task}"
)
research_result = self.researcher.think([research_msg])
self._log("技术调研", research_result)
results["phases"].append({
"phase": "research", "output": research_result
})

# ========== 第二阶段：代码生成 ==========
print("\\n💻 第二阶段：代码生成")
code_msg = Message(
role="user", sender="Coordinator",
content=f"""请根据以下需求和调研结果编写代码：

## 需求
{task}

## 技术调研结果
{research_result}

请直接输出完整可运行的 Python 代码。"""
)
code_result = self.coder.think([code_msg])
self._log("代码生成", code_result)
results["phases"].append({
"phase": "code", "output": code_result
})

# ========== 第三阶段：代码审查（可循环） ==========
current_code = code_result
revision = 0

while revision <= max_revisions:
print(f"\\n🧪 第三阶段：代码审查（第 {revision + 1} 次）")
review_msg = Message(
role="user", sender="Coordinator",
content=f"请审查以下代码：\\n{current_code}"
)
review_result = self.reviewer.think([review_msg])
self._log(f"代码审查（第{revision+1}次）", review_result)

# 检查评分（简单解析）
score = self._extract_score(review_result)
print(f"\\n📊 当前评分：{score}/10")

if score >= 8 or revision == max_revisions:
results["phases"].append({
"phase": "review",
"output": review_result,
"score": score
})
break

# 分数不够，让程序员修改
print(f"\\n🔄 评分不足，进入修改…")
fix_msg = Message(
role="user", sender="Coordinator",
content=f"""请根据审查意见修改代码：

## 审查意见
{review_result}

## 原始代码
{current_code}

请输出修改后的完整代码。"""
)
current_code = self.coder.think([fix_msg])
self._log(f"代码修改（第{revision+1}次）", current_code)
revision += 1

results["final_code"] = current_code
results["revisions"] = revision + 1

return results

def _extract_score(self, review_text: str) –> int:
"""从审查结果中提取评分"""
for line in review_text.split("\\n"):
if "评分" in line and "/" in line:
try:
score = int(line.split(":")[–1].strip().split("/")[0])
return score
except (ValueError, IndexError):
pass
return 5 # 默认评分

4.4 主程序：启动你的 AI 团队

# main.py
from agents import researcher, coder, reviewer
from coordinator import Coordinator
import json

def main():
print("🚀 AI 研发小队已就绪！")
print("=" * 60)

# 初始化协调者
coordinator = Coordinator(researcher, coder, reviewer)

# 示例任务
task = "用 Python 写一个文件批量重命名工具，支持按日期、序号、正则匹配重命名，需要命令行参数解析和日志记录"

# 执行任务
results = coordinator.run(task, max_revisions=2)

# 输出最终结果
print("\\n" + "=" * 60)
print("✅ 任务完成！")
print("=" * 60)
print(f"\\n📊 总修改轮数：{results['revisions']}")
print(f"\\n💻 最终代码：\\n{results['final_code']}")

# 保存结果
with open("output.json", "w", encoding="utf-8") as f:
json.dump(results, f, ensure_ascii=False, indent=2)
print("\\n💾 完整结果已保存到 output.json")

if __name__ == "__main__":
main()

五、执行流程详解

审查员

程序员

研究员

协调者

用户

审查员

程序员

研究员

协调者

用户

#mermaid-svg-o0Z2HPjQMlCc3cKI{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-o0Z2HPjQMlCc3cKI .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-o0Z2HPjQMlCc3cKI .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-o0Z2HPjQMlCc3cKI .error-icon{fill:#552222;}#mermaid-svg-o0Z2HPjQMlCc3cKI .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-o0Z2HPjQMlCc3cKI .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-o0Z2HPjQMlCc3cKI .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-o0Z2HPjQMlCc3cKI .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-o0Z2HPjQMlCc3cKI .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-o0Z2HPjQMlCc3cKI .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-o0Z2HPjQMlCc3cKI .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-o0Z2HPjQMlCc3cKI .marker{fill:#333333;stroke:#333333;}#mermaid-svg-o0Z2HPjQMlCc3cKI .marker.cross{stroke:#333333;}#mermaid-svg-o0Z2HPjQMlCc3cKI svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-o0Z2HPjQMlCc3cKI p{margin:0;}#mermaid-svg-o0Z2HPjQMlCc3cKI .actor{stroke:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);fill:#ECECFF;}#mermaid-svg-o0Z2HPjQMlCc3cKI text.actor>tspan{fill:black;stroke:none;}#mermaid-svg-o0Z2HPjQMlCc3cKI .actor-line{stroke:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);}#mermaid-svg-o0Z2HPjQMlCc3cKI .innerArc{stroke-width:1.5;stroke-dasharray:none;}#mermaid-svg-o0Z2HPjQMlCc3cKI .messageLine0{stroke-width:1.5;stroke-dasharray:none;stroke:#333;}#mermaid-svg-o0Z2HPjQMlCc3cKI .messageLine1{stroke-width:1.5;stroke-dasharray:2,2;stroke:#333;}#mermaid-svg-o0Z2HPjQMlCc3cKI #arrowhead path{fill:#333;stroke:#333;}#mermaid-svg-o0Z2HPjQMlCc3cKI .sequenceNumber{fill:white;}#mermaid-svg-o0Z2HPjQMlCc3cKI #sequencenumber{fill:#333;}#mermaid-svg-o0Z2HPjQMlCc3cKI #crosshead path{fill:#333;stroke:#333;}#mermaid-svg-o0Z2HPjQMlCc3cKI .messageText{fill:#333;stroke:none;}#mermaid-svg-o0Z2HPjQMlCc3cKI .labelBox{stroke:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);fill:#ECECFF;}#mermaid-svg-o0Z2HPjQMlCc3cKI .labelText,#mermaid-svg-o0Z2HPjQMlCc3cKI .labelText>tspan{fill:black;stroke:none;}#mermaid-svg-o0Z2HPjQMlCc3cKI .loopText,#mermaid-svg-o0Z2HPjQMlCc3cKI .loopText>tspan{fill:black;stroke:none;}#mermaid-svg-o0Z2HPjQMlCc3cKI .loopLine{stroke-width:2px;stroke-dasharray:2,2;stroke:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);fill:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);}#mermaid-svg-o0Z2HPjQMlCc3cKI .note{stroke:#aaaa33;fill:#fff5ad;}#mermaid-svg-o0Z2HPjQMlCc3cKI .noteText,#mermaid-svg-o0Z2HPjQMlCc3cKI .noteText>tspan{fill:black;stroke:none;}#mermaid-svg-o0Z2HPjQMlCc3cKI .activation0{fill:#f4f4f4;stroke:#666;}#mermaid-svg-o0Z2HPjQMlCc3cKI .activation1{fill:#f4f4f4;stroke:#666;}#mermaid-svg-o0Z2HPjQMlCc3cKI .activation2{fill:#f4f4f4;stroke:#666;}#mermaid-svg-o0Z2HPjQMlCc3cKI .actorPopupMenu{position:absolute;}#mermaid-svg-o0Z2HPjQMlCc3cKI .actorPopupMenuPanel{position:absolute;fill:#ECECFF;box-shadow:0px 8px 16px 0px rgba(0,0,0,0.2);filter:drop-shadow(3px 5px 2px rgb(0 0 0 / 0.4));}#mermaid-svg-o0Z2HPjQMlCc3cKI .actor-man line{stroke:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);fill:#ECECFF;}#mermaid-svg-o0Z2HPjQMlCc3cKI .actor-man circle,#mermaid-svg-o0Z2HPjQMlCc3cKI line{stroke:hsl(259.6261682243, 59.7765363128%, 87.9019607843%);fill:#ECECFF;stroke-width:2px;}#mermaid-svg-o0Z2HPjQMlCc3cKI :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

第一阶段：调研

第二阶段：编码

第三阶段：审查

评分不足，要求修改

提交任务需求

需求分析 + 技术方案

返回调研报告

调研结果 + 编码需求

返回代码

提交代码审查

评分 6/10 + 修改建议

审查意见 + 修改要求

返回修改后代码

再次审查

评分 9/10 ✓

交付最终代码 + 报告

六、扩展：加入工具能力

让 Agent 不只是"说话"，还能"做事"：

# tools.py
import os
import subprocess

def write_file(path: str, content: str) –> str:
"""写入文件"""
os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
with open(path, "w", encoding="utf-8") as f:
f.write(content)
return f"文件已写入: {path}"

def run_python(filepath: str) –> str:
"""运行 Python 文件并返回结果"""
try:
result = subprocess.run(
["python", filepath],
capture_output=True, text=True, timeout=30
)
output = result.stdout
if result.returncode != 0:
output += f"\\n❌ 错误:\\n{result.stderr}"
return output
except subprocess.TimeoutExpired:
return "执行超时（30秒）"

# 将工具注册到 Agent
AVAILABLE_TOOLS = {
"write_file": {
"func": write_file,
"description": "将内容写入指定文件",
"params": ["path", "content"]
},
"run_python": {
"func": run_python,
"description": "运行 Python 文件",
"params": ["filepath"]
}
}

七、性能对比

在我的测试环境（i7-13700K + 32GB 内存，纯 CPU 模式）下的表现：

#mermaid-svg-KqkvYDLUjZtAiopO{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-KqkvYDLUjZtAiopO .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-KqkvYDLUjZtAiopO .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-KqkvYDLUjZtAiopO .error-icon{fill:#552222;}#mermaid-svg-KqkvYDLUjZtAiopO .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-KqkvYDLUjZtAiopO .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-KqkvYDLUjZtAiopO .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-KqkvYDLUjZtAiopO .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-KqkvYDLUjZtAiopO .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-KqkvYDLUjZtAiopO .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-KqkvYDLUjZtAiopO .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-KqkvYDLUjZtAiopO .marker{fill:#333333;stroke:#333333;}#mermaid-svg-KqkvYDLUjZtAiopO .marker.cross{stroke:#333333;}#mermaid-svg-KqkvYDLUjZtAiopO svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-KqkvYDLUjZtAiopO p{margin:0;}#mermaid-svg-KqkvYDLUjZtAiopO .pieCircle{stroke:#000000;stroke-width:2px;opacity:0.7;}#mermaid-svg-KqkvYDLUjZtAiopO .pieOuterCircle{stroke:#000000;stroke-width:1px;fill:none;}#mermaid-svg-KqkvYDLUjZtAiopO .pieTitleText{text-anchor:middle;font-size:25px;fill:#000000;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}#mermaid-svg-KqkvYDLUjZtAiopO .slice{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;fill:#000000;font-size:17px;}#mermaid-svg-KqkvYDLUjZtAiopO .legend text{fill:#000000;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:17px;}#mermaid-svg-KqkvYDLUjZtAiopO :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

40%

25%

20%

12%

3%

多 Agent 系统各阶段耗时占比

技术调研（Researcher）

代码生成（Coder）

代码审查（Reviewer）

修改迭代（Revision）

协调调度开销

如有 GPU，速度可提升 3-5 倍。

八、注意事项

总结

#mermaid-svg-WI2uzVUKwBf6kWJU{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-WI2uzVUKwBf6kWJU .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-WI2uzVUKwBf6kWJU .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-WI2uzVUKwBf6kWJU .error-icon{fill:#552222;}#mermaid-svg-WI2uzVUKwBf6kWJU .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-WI2uzVUKwBf6kWJU .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-WI2uzVUKwBf6kWJU .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-WI2uzVUKwBf6kWJU .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-WI2uzVUKwBf6kWJU .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-WI2uzVUKwBf6kWJU .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-WI2uzVUKwBf6kWJU .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-WI2uzVUKwBf6kWJU .marker{fill:#333333;stroke:#333333;}#mermaid-svg-WI2uzVUKwBf6kWJU .marker.cross{stroke:#333333;}#mermaid-svg-WI2uzVUKwBf6kWJU svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-WI2uzVUKwBf6kWJU p{margin:0;}#mermaid-svg-WI2uzVUKwBf6kWJU .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-WI2uzVUKwBf6kWJU .cluster-label text{fill:#333;}#mermaid-svg-WI2uzVUKwBf6kWJU .cluster-label span{color:#333;}#mermaid-svg-WI2uzVUKwBf6kWJU .cluster-label span p{background-color:transparent;}#mermaid-svg-WI2uzVUKwBf6kWJU .label text,#mermaid-svg-WI2uzVUKwBf6kWJU span{fill:#333;color:#333;}#mermaid-svg-WI2uzVUKwBf6kWJU .node rect,#mermaid-svg-WI2uzVUKwBf6kWJU .node circle,#mermaid-svg-WI2uzVUKwBf6kWJU .node ellipse,#mermaid-svg-WI2uzVUKwBf6kWJU .node polygon,#mermaid-svg-WI2uzVUKwBf6kWJU .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-WI2uzVUKwBf6kWJU .rough-node .label text,#mermaid-svg-WI2uzVUKwBf6kWJU .node .label text,#mermaid-svg-WI2uzVUKwBf6kWJU .image-shape .label,#mermaid-svg-WI2uzVUKwBf6kWJU .icon-shape .label{text-anchor:middle;}#mermaid-svg-WI2uzVUKwBf6kWJU .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-WI2uzVUKwBf6kWJU .rough-node .label,#mermaid-svg-WI2uzVUKwBf6kWJU .node .label,#mermaid-svg-WI2uzVUKwBf6kWJU .image-shape .label,#mermaid-svg-WI2uzVUKwBf6kWJU .icon-shape .label{text-align:center;}#mermaid-svg-WI2uzVUKwBf6kWJU .node.clickable{cursor:pointer;}#mermaid-svg-WI2uzVUKwBf6kWJU .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-WI2uzVUKwBf6kWJU .arrowheadPath{fill:#333333;}#mermaid-svg-WI2uzVUKwBf6kWJU .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-WI2uzVUKwBf6kWJU .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-WI2uzVUKwBf6kWJU .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-WI2uzVUKwBf6kWJU .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-WI2uzVUKwBf6kWJU .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-WI2uzVUKwBf6kWJU .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-WI2uzVUKwBf6kWJU .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-WI2uzVUKwBf6kWJU .cluster text{fill:#333;}#mermaid-svg-WI2uzVUKwBf6kWJU .cluster span{color:#333;}#mermaid-svg-WI2uzVUKwBf6kWJU div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-WI2uzVUKwBf6kWJU .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-WI2uzVUKwBf6kWJU rect.text{fill:none;stroke-width:0;}#mermaid-svg-WI2uzVUKwBf6kWJU .icon-shape,#mermaid-svg-WI2uzVUKwBf6kWJU .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-WI2uzVUKwBf6kWJU .icon-shape p,#mermaid-svg-WI2uzVUKwBf6kWJU .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-WI2uzVUKwBf6kWJU .icon-shape .label rect,#mermaid-svg-WI2uzVUKwBf6kWJU .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-WI2uzVUKwBf6kWJU .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-WI2uzVUKwBf6kWJU .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-WI2uzVUKwBf6kWJU :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

能力有限

分工明确

互相审查

并行处理

单 Agent

多 Agent 协作

更高质量输出

更少 Bug

更快交付

多智能体系统的核心理念：不是让一个 AI 做所有事，而是让多个专精 AI 各司其职、互相补位。

就像真正的团队一样——没有人是全能的，但一个好的团队可以。