AI Agent Skills 开发者指南 2026 版

AI Agent Skills 是构建 AI Agent 技能的框架。从 Skill 定义到部署。2026 开发者必读指南。

  • ⭐ 3400
  • Python
  • YAML
  • MIT
  • 更新于 2026-05-18

什么是 AI Agent Skills? #

AI Agent Skills 是构建 AI Agent 技能的框架。把「能力」模块化。从定义 Skill 到部署运行。

目录结构 #

my_skill/
├── SKILL.md       # 文档
├── skill.py       # 实现
├── tests/         # 单元测试
└── examples/      # 使用示例

SKILL.md 格式 #

---
name: 天气查询
description: 查询全球城市天气
tags: ["weather", "api", "daily"]
version: 1.0.0
author: your-name
date: 2026-05-18
---

## 功能描述

查询指定城市的当前天气。

## 使用方法

```python
from skill import WeatherSkill
skill = WeatherSkill()
result = skill.run("北京")

返回格式 #

{"temp": 22, "condition": "晴天", "humidity": 65}

## 创建 Skill

### 使用 Hermes

```bash
hermes skills create weather-query

手动创建 #

mkdir -p weather-query
touch weather-query/SKILL.md
touch weather-query/skill.py

Skill 实现 #

Python 示例 #

import requests

class WeatherSkill:
    name = "天气查询"
    description = "查询城市天气"
    
    def __init__(self, api_key: str):
        self.api_key = api_key
    
    def run(self, city: str) -> dict:
        url = f"https://api.weather.com/v1/weather"
        params = {"city": city, "key": self.api_key}
        response = requests.get(url, params=params)
        return response.json()

集成记忆 #

from agent_memory import MemoryMixin

class WeatherSkill(MemoryMixin):
    def run(self, city: str) -> dict:
        key = f"weather:{city}"
        cached = self.cache.get(key)
        if cached:
            return cached
        
        result = self._fetch_weather(city)
        self.cache.set(key, result, ttl=3600)
        return result

部署流程 #

1. 编写代码 #

# skill.py
from base_skill import BaseSkill

class MySkill(BaseSkill):
    def execute(self, params: dict) -> dict:
        return {"result": "hello"}

2. 编写测试 #

def test_skill():
    skill = MySkill()
    result = skill.execute({"name": "test"})
    assert result["result"] == "hello"

3. 集成到 Agent #

from agent import Agent
from skills import MySkill

agent = Agent(skills=[MySkill])
response = agent.run("调用我的 Skill")

高级功能 #

并行执行 #

from concurrent.futures import ThreadPoolExecutor

skills = [Skill1(), Skill2(), Skill3()]
with ThreadPoolExecutor() as executor:
    results = list(executor.map(lambda s: s.run(data), skills))

错误处理 #

from skill.exceptions import SkillError

try:
    result = skill.run(data)
except SkillError as e:
    result = skill.fallback(data)

版本管理 #

# 版本化
git tag v1.0.0

# 更新
pip install my-skill==1.1.0

常见问题 #

Q: Skill 会超时吗?

答:支持超时控制。skill.run(data, timeout=30)

Q: 如何分享 Skill?

答:发布到 PyPI,或分享 Git 仓库链接。

Q: 支持异步调用吗?

答:支持。async def run()await skill.run()

总结 #

AI Agent Skills 把「能力」变成「可插即拔」的组件。代码即技能。


参考:ai-agent-skills.org 文档 更新:2026-05-18

💬 留言讨论