完整、可运行思路 + 示例代码。是目前社区里最常见、也最推荐的做法（YAML + Python Factory 模式）。

---

一、目录结构示例（推荐）

my_crew/
├── agents/
│   ├── researcher.yaml
│   ├── writer.yaml
│   └── critic.yaml
├── tools/
│   └── search_tool.py
├── config_loader.py      # Agent 工厂
└── main.py               # 启动 Crew

---

二、Agent 配置文件（YAML）

agents/researcher.yaml

role: Senior Researcher
goal: Find accurate and up-to-date information
backstory: >
  You are an expert researcher with a PhD in AI.
  You are meticulous and always cite sources.
tools:
  - name: SearchTool
    module: tools.search_tool
    class: SearchTool
verbose: true
allow_delegation: false

agents/writer.yaml

role: Content Writer
goal: Write engaging blog posts
backstory: >
  You are a professional copywriter who specializes in AI topics.
tools: []
verbose: true

✅ 说明

• tools支持动态导入 Python 类

• backstory对应你说的 soul.md内容

• 每个 agent 一个文件，易维护

---

三、Tool 示例（Skill 的实现）

tools/search_tool.py

from crewai.tools import BaseTool

class SearchTool(BaseTool):
    name = "SearchTool"
    description = "Search the web for information"

    def _run(self, query: str) -> str:
        return f"Search results for: {query}"

---

四、Agent 工厂（核心：动态加载）

config_loader.py

import yaml
from crewai import Agent
from importlib import import_module

def load_tool(tool_config):
    """动态加载 Tool"""
    module = import_module(tool_config["module"])
    tool_class = getattr(module, tool_config["class"])
    return tool_class()

def load_agent_from_yaml(yaml_path: str) -> Agent:
    with open(yaml_path, "r", encoding="utf-8") as f:
        config = yaml.safe_load(f)

    tools = []
    for tool_cfg in config.get("tools", []):
        tools.append(load_tool(tool_cfg))

    agent = Agent(
        role=config["role"],
        goal=config["goal"],
        backstory=config["backstory"],
        tools=tools,
        verbose=config.get("verbose", False),
        allow_delegation=config.get("allow_delegation", True),
    )
    return agent

def load_all_agents(agents_dir: str) -> list[Agent]:
    import os
    agents = []
    for file in os.listdir(agents_dir):
        if file.endswith(".yaml"):
            path = os.path.join(agents_dir, file)
            agents.append(load_agent_from_yaml(path))
    return agents

✅ 这里实现了：

• YAML → Python 对象

• Tool 的反射式加载

• 多 Agent 批量初始化

---

五、在 main.py 中启动 Crew

from crewai import Crew, Task
from config_loader import load_all_agents

agents = load_all_agents("agents")

research_task = Task(
    description="Research the latest trends in CrewAI",
    agent=agents[0]  # researcher
)

write_task = Task(
    description="Write a blog post based on research",
    agent=agents[1]  # writer
)

crew = Crew(
    agents=agents,
    tasks=[research_task, write_task],
    verbose=True
)

result = crew.kickoff()
print(result)

---

六、进阶玩法（可选）

✅ 1. 支持 soul.md / skill.md（社区风格）

你可以把：

backstory_file: souls/researcher.md
skills_file: skills/researcher.md

然后在 loader 中：

with open(config["backstory_file"]) as f:
    backstory = f.read()

---

✅ 2. 支持 JSON / TOML

CrewAI 不关心格式，只要你最终构造出 Agent()即可。

---

七、总结一句话

CrewAI 本身不“存储” agents，而是通过 Python 对象 + 你自己的配置加载机制来实现动态管理。