Agent

参考：

• Build a Coding Agent from Scratch: The Complete Python Tutorial - Sid Bharath

1. system_prompt

   1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

   SYSTEM_PROMPT = """You are a helpful coding agent that assists with programming tasks and file operations. When responding to requests: 1. Analyze what the user needs 2. Use the minimum number of tools necessary to accomplish the task 3. After using tools, provide a concise summary of what was done IMPORTANT: Once you've completed the requested task, STOP and provide your final response. Do not continue creating additional files or performing extra actions unless specifically asked. Examples of good behavior: - User: "Create a file that adds numbers" → Create ONE file, then summarize - User: "Create files for add and subtract" → Create ONLY those two files, then summarize - User: "Create math operation files" → Ask for clarification on which operations, or create a reasonable set and stop After receiving tool results: - If the task is complete, provide a final summary - Only continue with more tools if the original request is not yet fulfilled - Do not interpret successful tool execution as a request to do more Be concise and efficient. Complete the requested task and stop."""

Reason –> Action –> Observe

• Reason: 是否/使用哪些工具
• Action: 调用工具
• Observe: 观察工具调用结果

两种调用方式

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

def auto_tool_use(self, user: str, max_tool_auto_call: int):

    n = 0
    while n < max_tool_auto_call:
        # 1) 整理 messages
        messages = self.messages[max_history_count:]
        # 2) llm 调用 :::: Reason
        ret = self.llm.invode(user)

        n += 1
        # 3) 判断是否需要 tool 调用
        if "tool_use" not in ret:
            msg = ret.content
            messages.append({"role": "user", "content": user})
            messages.append({"role": "assistant", "content": msg})
            break

        # 4) 执行 tool 调用 :::: Action
        tool_calls = ret.tool_calls
        tool_results: list[dict] = self._execute__tools(tool_calls)
        # 5) 保存 tool 调用结果 :::: Observe
        self.messages.extend(
            [
                {
                    "role": "user",
                    # 注意：这里content不再是str, 而是一个 list[dict]
                    "content": [
                        {
                            "type": "tool_result",
                            "tool_use_id": t["tool_use_id"],
                            "content": t["content"],
                        }
                        for t in tool_results
                    ],
                }
            ]
        )
        # 6) 下一轮 Reasoning --> Action --> Oberve
        #    根据 tool 调用的结果判断是否已经结果问题，如果没有结果，LLM 根据需要再次触发 tool 调用
    return messages

参考：

• A Guide to Context Engineering: Setting Agents Up for Success - Sid Bharath

Claude Code 包括了三层 context (three-layout context)

1. 第一层： Cluade.md

   • persistent layer

     • 长期记忆
   • 在初始化一个项目任务时，Cluade 会先扫描项目的重要文件（requirements.txt, README,项目结构等等），然后创建 Cluade.md

   • 包括内容：

     • projtect convention 项目风格
     • commit 规范
     • architecture decisions
     • coding standards
     • 已经你个人规定的其他要求

2. 第二层： dynamic layer

   • 根据用户的 query 命令，去获取项目中的相关内容来增强 context

   • 比如：用户提取修复 auth bug, 它会

     1. 查找和认证相关的文件和代码，理解代码逻辑，查找相关 codebase,比如 middleware, config, tests
     2. 扩充 context
     3. 制定修复 bug 的方案
     4. 让用户确定是否采用

3. 第三层：conversation

   • 对话本身

     1. user query
     2. tools used
     3. content retrieved

• todolist 作为 middleware 实现，而不是 tool

  • https://claude.ai/share/a7ddb21d-e9c2-4276-8af9-a48190e39585

  • 如果使用 tool 实现，潜在的问题

    1. ❌ 模型可能**忘记**使用 todo 工具
    2. ❌ 模型需要**主动决定**何时添加任务
    3. ❌ 在 prompt 中占用 token（所有工具都会被列出）
    4. ❌ 没有强制规划流程

  • 使用 middleware 的好处

    1. 模型无感知
    2. 强制调用

• 使用 middleware 实现，底层调用逻辑顺序

  1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39

  from langchain.agents import create_agent from langchain.agents.middleware import TodoListMiddleware # ✅ 用 Middleware 实现（自动执行） agent = create_agent( model="gpt-4o", tools=[search_web, send_email], # 不需要添加 todo 工具 middleware=[TodoListMiddleware()] ) agent.invoke("Help me plan my day and send emails") ``` ,**优势：** 1. ✅ **自动执行**：模型在思考前会自动更新待办 2. ✅ **不占用工具槽**：不会在模型的工具列表中 3. ✅ **强制规划**：系统提示会要求模型维护任务列表 4. ✅ **透明管理**：模型无需关心实现细节 ,**实际执行：** ``` [before_model 钩子自动触发] System: "Use write_todos to plan your tasks..." AI calls: write_todos([ {content: "Search for weather", status: "pending"}, {content: "Draft email", status: "pending"}, {content: "Send email", status: "pending"} ]) AI: Let me start by searching the weather... [执行工具] [after_model 钩子自动更新状态] write_todos([ {content: "Search for weather", status: "completed"}, {content: "Draft email", status: "in_progress"}, ... ])