AI 智能体构建器 (AI Agent Builder)

设计并构建具有工具、记忆和多步推理能力的 AI 智能体。涵盖基于 n8n 5000+ AI 工作流模板的 ChatGPT、Claude、Gemini 集成模式。

概述 (Overview)

本技能涵盖：

• AI 智能体架构设计
• 工具/函数调用模式
• 记忆与上下文管理
• 多步推理工作流
• 平台集成（Slack、Telegram、Web）

AI 智能体架构 (AI Agent Architecture)

核心组件 (Core Components)

┌─────────────────────────────────────────────────────────────────┐
│                      AI 智能体架构 (AI AGENT ARCHITECTURE)        │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  ┌─────────────┐     ┌─────────────┐     ┌─────────────┐        │
│  │   输入       │────▶│   智能体     │────▶│   输出      │        │
│  │  (查询)      │     │   (LLM)     │     │  (响应)     │        │
│  └─────────────┘     └──────┬──────┘     └─────────────┘        │
│                             │                                   │
│         ┌───────────────────┼───────────────────┐               │
│         │                   │                   │               │
│         ▼                   ▼                   ▼               │
│  ┌─────────────┐     ┌─────────────┐     ┌─────────────┐        │
│  │   工具       │     │   记忆       │     │   知识库     │       │
│  │  (函数)      │     │  (上下文)    │     │   (RAG)     │        │
│  └─────────────┘     └─────────────┘     └─────────────┘        │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

智能体类型 (Agent Types)

agent_types:
  reactive_agent:
    description: "单轮响应，无记忆 (Single-turn response, no memory)"
    use_case: 简单问答, 分类 (simple_qa, classification)
    complexity: 低 (low)
    
  conversational_agent:
    description: "多轮对话，具有对话记忆 (Multi-turn with conversation memory)"
    use_case: 聊天机器人, 客服支持 (chatbots, support)
    complexity: 中 (medium)
    
  tool_using_agent:
    description: "可调用外部工具/API (Can call external tools/APIs)"
    use_case: 数据查询, 执行操作 (data_lookup, actions)
    complexity: 中 (medium)
    
  reasoning_agent:
    description: "多步规划与执行 (Multi-step planning and execution)"
    use_case: 复杂任务, 研究分析 (complex_tasks, research)
    complexity: 高 (high)
    
  multi_agent:
    description: "多个专业智能体协作 (Multiple specialized agents collaborating)"
    use_case: 复杂工作流 (complex_workflows)
    complexity: 非常高 (very_high)

工具调用模式 (Tool Calling Pattern)

工具定义 (Tool Definition)

tool_definition:
  name: "get_weather"
  description: "获取某个位置的当前天气 (Get current weather for a location)"
  parameters:
    type: object
    properties:
      location:
        type: string
        description: "城市名或坐标 (City name or coordinates)"
      units:
        type: string
        enum: ["celsius", "fahrenheit"]
        default: "celsius"
    required: ["location"]
    
  implementation:
    type: api_call
    endpoint: "https://api.weather.com/v1/current"
    method: GET
    params:
      q: "{location}"
      units: "{units}"

常见工具类别 (Common Tool Categories)

tool_categories:
  data_retrieval: # 数据检索
    - web_search: 搜索互联网 (search the internet)
    - database_query: 查询 SQL/NoSQL (query SQL/NoSQL)
    - api_lookup: 调用外部 API (call external APIs)
    - file_read: 读取文档 (read documents)
    
  actions: # 操作
    - send_email: 发送邮件 (send emails)
    - create_calendar: 安排日程 (schedule events)
    - update_crm: 修改 CRM 记录 (modify CRM records)
    - post_slack: 发送 Slack 消息 (send Slack messages)
    
  computation: # 计算
    - calculator: 数学运算 (math operations)
    - code_interpreter: 运行 Python (run Python)
    - data_analysis: 分析数据集 (analyze datasets)
    
  generation: # 生成
    - image_generation: 创建图像 (create images)
    - document_creation: 生成文档 (generate docs)
    - chart_creation: 创建可视化 (create visualizations)

n8n 工具集成 (n8n Tool Integration)

n8n_agent_workflow:
  nodes:
    - trigger:
        type: webhook
        path: "/ai-agent"
        
    - ai_agent:
        type: "@n8n/n8n-nodes-langchain.agent"
        model: openai_gpt4
        system_prompt: |
          你是一个乐于助人的助手，可以：
          (You are a helpful assistant that can:)
          1. 搜索网络获取信息 (Search the web for information)
          2. 查询我们的客户数据库 (Query our customer database)
          3. 代表用户发送邮件 (Send emails on behalf of the user)
          
        tools:
          - web_search
          - database_query
          - send_email
          
    - respond:
        type: respond_to_webhook
        data: "{{ $json.output }}"

记忆模式 (Memory Patterns)

记忆类型 (Memory Types)

memory_types:
  buffer_memory:
    description: "存储最近 N 条消息 (Store last N messages)"
    implementation: |
      messages = []
      def add_message(role, content):
          messages.append({"role": role, "content": content})
          if len(messages) > MAX_MESSAGES:
              messages.pop(0)
    use_case: 简单聊天机器人 (simple_chatbots)
    
  summary_memory:
    description: "定期总结对话内容 (Summarize conversation periodically)"
    implementation: |
      When messages > threshold:
          summary = llm.summarize(messages[:-5])
          messages = [summary_message] + messages[-5:]
    use_case: 长对话 (long_conversations)
    
  vector_memory:
    description: "存储在向量数据库中用于语义检索 (Store in vector DB for semantic retrieval)"
    implementation: |
      # 存储 (Store)
      embedding = embed(message)
      vector_db.insert(embedding, message)
      
      # 检索 (Retrieve)
      relevant = vector_db.search(query_embedding, k=5)
    use_case: 知识检索 (knowledge_retrieval)
    
  entity_memory:
    description: "跟踪对话中提到的实体 (Track entities mentioned in conversation)"
    implementation: |
      entities = {}
      def update_entities(message):
          extracted = llm.extract_entities(message)
          entities.update(extracted)
    use_case: 个性化助手 (personalized_assistants)

上下文窗口管理 (Context Window Management)

context_management:
  strategies:
    sliding_window:
      keep: 最近 N 条消息 (last_n_messages)
      n: 10
      
    relevance_based:
      method: 嵌入并排序 (embed_and_rank)
      keep: 保留最相关的 K 条 (top_k_relevant)
      k: 5
      
    hierarchical:
      levels:
        - immediate: 最近 3 条消息 (last_3_messages)
        - recent: 最近 10 条消息的摘要 (summary_of_last_10)
        - long_term: 全部消息的关键事实 (key_facts_from_all)
        
  token_budget:
    total: 8000
    system_prompt: 1000
    tools: 1000
    memory: 4000
    current_query: 1000
    response: 1000

多步推理 (Multi-Step Reasoning)

ReAct 模式 (ReAct Pattern)

思考 (Thought): 我需要查找关于 X 的信息 (I need to find information about X)
行动 (Action): web_search("X")
观察 (Observation): [搜索结果 (search results)]
思考 (Thought): 根据结果，我还应该检查 Y (Based on the results, I should also check Y)
行动 (Action): database_query("SELECT * FROM Y")
观察 (Observation): [数据库结果 (database results)]
思考 (Thought): 现在我有足够的信息来回答了 (Now I have enough information to answer)
行动 (Action): respond("基于 X 和 Y 的最终答案 (Final answer based on X and Y)")

规划型智能体 (Planning Agent)

planning_workflow:
  step_1_plan:
    prompt: |
      任务: {user_request}
      
      创建完成此任务的分步计划。
      每一步都应具体且可操作。
      (Create a step-by-step plan to complete this task.
      Each step should be specific and actionable.)
      
    output: 编号步骤列表 (numbered_steps)
    
  step_2_execute:
    for_each: 步骤 (step)
    actions:
      - 执行步骤 (execute_step)
      - 验证结果 (validate_result)
      - 必要时调整 (adjust_if_needed)
      
  step_3_synthesize:
    prompt: |
      已完成的步骤: {executed_steps}
      结果: {results}
      
      为用户综合出最终回复。
      (Synthesize a final response for the user.)

平台集成 (Platform Integrations)

Slack 机器人智能体 (Slack Bot Agent)

slack_agent:
  trigger: slack_message
  
  workflow:
    1. receive_message:
        extract: [user, channel, text, thread_ts]
        
    2. get_context:
        if: thread_ts
        action: 获取线程历史 (fetch_thread_history)
        
    3. process_with_agent:
        model: gpt-4
        system: "你是一个乐于助人的 Slack 助手 (You are a helpful Slack assistant)"
        tools: [web_search, jira_lookup, calendar_check]
        
    4. respond:
        action: 发送至 Slack (post_to_slack)
        channel: "{channel}"
        thread_ts: "{thread_ts}"
        text: "{agent_response}"

Telegram 机器人智能体 (Telegram Bot Agent)

telegram_agent:
  trigger: telegram_message
  
  handlers:
    text_message:
      - 提取文本 (extract_text)
      - 使用 AI 处理 (process_with_ai)
      - 发送回复 (send_response)
      
    voice_message:
      - 使用 Whisper 转写 (transcribe_with_whisper)
      - 使用 AI 处理 (process_with_ai)
      - 发送文本或语音回复 (send_text_or_voice_response)
      
    image:
      - 使用视觉模型分析 (analyze_with_vision)
      - 使用 AI 处理 (process_with_ai)
      - 发送回复 (send_response)
      
    document:
      - 提取内容 (extract_content)
      - 使用 AI 处理 (process_with_ai)
      - 发送回复 (send_response)

Web 聊天界面 (Web Chat Interface)

web_chat_agent:
  frontend:
    type: react_component
    features:
      - 消息输入 (message_input)
      - 消息历史 (message_history)
      - 输入指示器 (typing_indicator)
      - 文件上传 (file_upload)
      
  backend:
    endpoint: "/api/chat"
    method: POST
    streaming: true
    
  session_management:
    method: jwt_token
    storage: redis
    ttl: 24 小时 (24_hours)

智能体模板 (Agent Templates)

客服支持智能体 (Customer Support Agent)

support_agent:
  name: "客服支持 AI (Customer Support AI)"
  model: gpt-4
  
  system_prompt: |
    你是 {company} 的客服支持智能体。
    (You are a customer support agent for {company}.)
    
    准则 (Guidelines):
    - 乐于助人、专业且富有同理心 (Be helpful, professional, and empathetic)
    - 使用知识库回答问题 (Use the knowledge base to answer questions)
    - 如果无法帮助，转接至人工 (Escalate to human if you cannot help)
    - 绝不捏造信息 (Never make up information)
    
    可用操作 (Available actions):
    - 搜索知识库 (Search knowledge base)
    - 查找客户账户 (Look up customer account)
    - 创建支持工单 (Create support ticket)
    - 转接人工客服 (Escalate to human agent)
    
  tools:
    - knowledge_search:
        description: "搜索帮助文章 (Search help articles)"
    - customer_lookup:
        description: "获取客户账户信息 (Get customer account info)"
    - create_ticket:
        description: "创建支持工单 (Create support ticket)"
    - escalate:
        description: "转接至人工客服 (Transfer to human agent)"
        
  escalation_triggers:
    - sentiment: 非常消极 (very_negative)
    - keywords: ["找人工客服", "经理", "投诉"] (["speak to human", "manager", "complaint"])
    - failed_resolution: 3 次尝试后 (after_3_attempts)

研究智能体 (Research Agent)

research_agent:
  name: "研究助手 (Research Assistant)"
  model: claude-opus
  
  system_prompt: |
    你是一个研究助手，帮助用户查找并综合任何主题的信息。
    (You are a research assistant that helps users find 
    and synthesize information on any topic.)
    
    流程 (Process):
    1. 理解研究问题 (Understand the research question)
    2. 规划搜索策略 (Plan search strategy)
    3. 跨多个来源执行搜索 (Execute searches across multiple sources)
    4. 综合研究结果 (Synthesize findings)
    5. 附带引用呈现结果 (Present with citations)
    
  tools:
    - web_search
    - academic_search
    - news_search
    - wikipedia_lookup
    
  output_format:
    - 执行摘要 (executive_summary)
    - 关键发现 (key_findings)
    - 详细分析 (detailed_analysis)
    - 引用来源 (sources_cited)

输出示例 (Output Example)

请求 (Request): “为预约安排设计一个 AI 智能体 (Design an AI agent for appointment scheduling)”

输出 (Output):

# 预约安排智能体设计 (Appointment Scheduling Agent Design)

## 架构 (Architecture)

```yaml
agent:
  name: "安排助手 (Scheduling Assistant)"
  model: gpt-4
  
  capabilities:
    - 检查日历可用性 (Check calendar availability)
    - 预约 (Book appointments)
    - 发送确认 (Send confirmations)
    - 处理改期 (Handle rescheduling)
    - 解答常见问题 (Answer FAQs)

所需工具 (Tools Required)

1. calendar_check
   • 输入 (Input): 日期范围、时长 (date_range, duration)
   • 输出 (Output): 可用时段列表 (available_slots[])
2. book_appointment
   • 输入 (Input): 日期时间、客户信息、服务类型 (datetime, customer_info, service_type)
   • 输出 (Output): 确认 ID (confirmation_id)
3. send_notification
   • 输入 (Input): 类型、接收人、详情 (type, recipient, details)
   • 输出 (Output): 成功/失败 (success/failure)
4. customer_lookup
   • 输入 (Input): 邮箱或电话 (email or phone)
   • 输出 (Output): 客户记录 (customer_record)

对话流程 (Conversation Flow)

用户 (User): "我想预约 (I'd like to book an appointment)"
智能体 (Agent): [思考 (Thought): 需要了解服务类型和首选时间 (Need to know service type and preferred time)]
智能体 (Agent): "我很乐意帮忙！您需要什么服务，有偏好的日期/时间吗？
        (I'd be happy to help! What service are you looking for, 
        and do you have a preferred date/time?)"

用户 (User): "理发，下周二下午 (Haircut, next Tuesday afternoon)"
智能体 (Agent): [行动 (Action): calendar_check(date="下周二 (next Tuesday)", time="下午 (afternoon)")]
智能体 (Agent): "下周二下午我有 2 点、3 点和 4 点半的时间。哪个最适合您？
        (I have availability at 2pm, 3pm, and 4:30pm next Tuesday. 
        Which works best for you?)"

用户 (User): "3 点吧 (3pm please)"
智能体 (Agent): [行动 (Action): book_appointment(datetime="周二 3 点 (Tuesday 3pm)", service="理发 (haircut)")]
智能体 (Agent): [行动 (Action): send_notification(type="确认 (confirmation)", ...)]
智能体 (Agent): "您已成功预约下周二下午 3 点的理发。我已将确认信息发送至您的邮箱。到时见！
        (You're all set for a haircut at 3pm next Tuesday. 
        I've sent a confirmation to your email. See you then!)"

n8n 实现 (n8n Implementation)

workflow:
  - webhook_trigger: /schedule-chat
  - ai_agent:
      tools: [calendar, booking, notification]
  - respond_to_user

---

*AI 智能体构建器技能 - Claude 办公技能套件的一部分 (AI Agent Builder Skill - Part of Claude Office Skills)*