🔄 Agent Skills自动化工作流实战

🎯 什么是Agent Skills工作流？

Agent Skills工作流是将多个Skill串联起来，形成端到端的自动化流程。核心概念：

• Trigger（触发器）：定时任务、Webhook、消息事件
• Pipeline（管线）：Skill按顺序执行，前一步的输出是后一步的输入
• Action（动作）：最终执行的操作，如发送消息、创建文件、调用API
• Delivery（投递）：结果通知，支持announce、webhook、channel等模式

🚀 实战案例1：每日新闻聚合

Step 1：创建Cron定时任务

{
  "name": "daily-ai-news",
  "schedule": { "kind": "cron", "expr": "0 8 * * *", "tz": "Asia/Shanghai" },
  "payload": {
    "kind": "agentTurn",
    "message": "执行每日AI新闻聚合任务：1)抓取OpenAI、Anthropic、HuggingFace的RSS 2)筛选重要新闻 3)生成中文摘要 4)发送到Discord #ai-news频道"
  },
  "delivery": { "mode": "announce", "channel": "discord" }
}

Step 2：RSS采集Skill

---
name: rss_collector
version: 1.0.0
description: 多源RSS聚合采集
tools:
  - name: collect
    description: 从多个RSS源采集新闻
    parameters:
      feeds:
        type: array
        description: RSS源URL列表
        required: true
      since_hours:
        type: number
        description: 仅获取N小时内的新文章
        default: 24
  - name: summarize
    description: AI生成中文摘要
    parameters:
      articles:
        type: array
        description: 文章列表
        required: true
      max_items:
        type: number
        description: 最多摘要数量
        default: 10
---

🚀 实战案例2：GitHub自动运营

多步骤Cron配置

# 早上9点 - 监控并回复Discussions
{
  "name": "github-morning-patrol",
  "schedule": { "kind": "cron", "expr": "0 9 * * *", "tz": "Asia/Shanghai" },
  "payload": {
    "kind": "agentTurn",
    "message": "检查GitHub Discussions新帖：1)用gh CLI列出过去1小时的新帖子 2)分析内容分类(question/bug/feature) 3)搜索知识库生成回复 4)对无法回答的打上needs-attention标签"
  },
  "delivery": { "mode": "none" },
  "failureAlert": { "after": 3, "channel": "discord" }
}

# 晚上6点 - 生成运营日报
{
  "name": "github-daily-report",
  "schedule": { "kind": "cron", "expr": "0 18 * * *", "tz": "Asia/Shanghai" },
  "payload": {
    "kind": "agentTurn",
    "message": "GitHub运营日报：1)统计今日回答数、新建issue数、star增长 2)分析热点讨论话题 3)生成Markdown报告保存到/var/www/miaoquai/github-report.html"
  }
}

🚀 实战案例3：多渠道智能客服

消息路由逻辑

---
name: customer_service_router
version: 2.0.0
description: 智能客服消息路由与回复
tools:
  - name: route_message
    description: 根据消息内容自动分类路由
    parameters:
      message:
        type: string
        description: 客户消息原文
        required: true
      channel:
        type: string
        description: 来源渠道
        required: true
    returns:
      type: object
      properties:
        category:
          type: string
          enum: [sales, support, billing, partnership, general]
        confidence:
          type: number
        suggested_response:
          type: string

  - name: auto_reply
    description: 基于知识库自动回复
    parameters:
      message_id:
        type: string
        description: 消息ID
        required: true
      response:
        type: string
        description: 回复内容
        required: true

  - name: escalate
    description: 升级到人工客服
    parameters:
      message_id:
        type: string
        required: true
      reason:
        type: string
        required: true
      priority:
        type: string
        enum: [low, medium, high, urgent]
        default: medium
---

# 路由规则
## 销售类 → sales_agent
关键词：价格、多少钱、付费、套餐、优惠、购买

## 技术支持 → support_agent
关键词：bug、报错、无法、错误、崩溃、安装

## 账单问题 → billing_agent
关键词：发票、退款、扣费、账单、续费

## 合作咨询 → business_agent
关键词：合作、商务、API、集成、企业版

## 默认 → general_agent
使用知识库搜索最佳匹配答案

📊 工作流编排最佳实践

# ✅ 好实践：独立执行，容错处理
for step in workflow_steps:
    try:
        result = execute_step(step)
        checkpoint(step.name, result)
    except Exception as e:
        log_error(step.name, e)
        if step.critical:
            alert_team(step.name, e)
        continue  # 非关键步骤跳过，不中断流程

# 幂等操作：重复执行不会产生副作用
def create_daily_report(date):
    existing = check_report_exists(date)
    if existing:
        return existing  # 已存在则返回
    return generate_and_save_report(date)

# 非幂等操作：需要去重
def send_notification(user, message):
    if already_sent(user.id, message.hash):
        return  # 避免重复发送
    deliver(user, message)

⚡ 高级模式：SubAgent并行编排

# 使用sessions_spawn并行执行多个子Agent
# 主Agent编排
{
  "plan": [
    { "agent": "news_collector", "task": "抓取AI新闻RSS", "mode": "run" },
    { "agent": "github_monitor", "task": "检查GitHub新讨论", "mode": "run" },
    { "agent": "competitor_watcher", "task": "监控竞品动态", "mode": "run" },
    { "agent": "report_generator", "task": "汇总所有数据生成日报", "mode": "run", "depends_on": ["news_collector", "github_monitor", "competitor_watcher"] }
  ]
}

# 子Agent调用示例
sessions_spawn(
    task="抓取以下RSS源的新闻：OpenAI Blog、Anthropic Blog、HuggingFace Blog。返回JSON格式结果。",
    mode="run",       # 一次性执行
    runtime="subagent",
    label="news_collector"
)