从0到1构建像Claude Code那样的Agent（一）：Agent循环

阿杜杜不是阿木木

683人浏览 · 2026-03-17 20:21:02

阿杜杜不是阿木木 · 2026-03-17 20:21:02 发布

文章目录

1.Debug位置

def post(
        self,
        path: str,
        *,
        cast_to: Type[ResponseT],
        body: Body | None = None,
        content: BinaryTypes | None = None,
        options: RequestOptions = {},
        files: RequestFiles | None = None,
        stream: bool = False,
        stream_cls: type[_StreamT] | None = None,
) -> ResponseT | _StreamT:
    if body is not None and content is not None:
        raise TypeError("Passing both `body` and `content` is not supported")
    if files is not None and content is not None:
        raise TypeError("Passing both `files` and `content` is not supported")
    if isinstance(body, bytes):
        warnings.warn(
            "Passing raw bytes as `body` is deprecated and will be removed in a future version. "
            "Please pass raw bytes via the `content` parameter instead.",
            DeprecationWarning,
            stacklevel=2,
        )
    print("assistant request:", json.dumps(body, indent=2, ensure_ascii=False))
    opts = FinalRequestOptions.construct(
        method="post", url=path, json_data=body, content=content, files=to_httpx_files(files), **options
    )
    return cast(ResponseT, self.request(cast_to, opts, stream=stream, stream_cls=stream_cls))

s01_agent_loop.py#73 #89

def agent_loop(messages: list):
    while True:
        response = client.messages.create(
            model=MODEL, system=SYSTEM, messages=messages,
            tools=TOOLS, max_tokens=8000,
        )
        print("assistant response:", response.to_json())
        # Append assistant turn
        messages.append({"role": "assistant", "content": response.content})
        # If the model didn't call a tool, we're done
        if response.stop_reason != "tool_use":
            return
        # Execute each tool call, collect results
        results = []
        for block in response.content:
            if block.type == "tool_use":
                print(f"\033[33m$ {block.input['command']}\033[0m")
                output = run_bash(block.input["command"])
                print(output[:200])
                results.append({"type": "tool_result", "tool_use_id": block.id,
                                "content": output})
        messages.append({"role": "user", "content": results})
        print("messages:", json.dumps(messages, indent=2, ensure_ascii=False))

2.Agent循环

Single-tool agent loop 单工具代理循环

The minimal agent kernel is a while loop + one tool
最小代理内核是一个 while 循环+一个工具

在这里插入图片描述

"One loop & Bash is all you need" -- 一个工具 + 一个循环 = 一个智能体。

3.问题

语言模型能推理代码, 但碰不到真实世界 – 不能读文件、跑测试、看报错。没有循环, 每次工具调用你都得手动把结果粘回去。你自己就是那个循环。

4.设计决策

为什么仅靠 Bash 就够了?

Bash 能读写文件、运行任意程序、在进程间传递数据、管理文件系统。任何额外的工具（read_file、write_file 等）都只是 bash 已有能力的子集。增加工具并不会解锁新能力，只会增加模型需要理解的接口。模型只需学习一个工具的 schema，实现代码不超过 100 行。这就是最小可行 agent：一个工具，一个循环。

我们本可以开始时使用更丰富的工具集（文件 I/O、HTTP、数据库），但这会掩盖核心见解：一个带有外壳的 LLM 已经是一个通用代理。从最简开始也使得每个后续版本实际添加的内容一目了然。

用递归进程创建实现子代理机制?

当 agent 执行 python v0.py "subtask" 时，它会创建一个全新的进程，拥有全新的 LLM 上下文。这个子进程实际上就是一个子代理：有自己的系统提示词、对话历史和任务焦点。子进程完成后，父进程通过 stdout 获取结果。这就是不依赖任何框架的子代理委派——纯粹的 Unix 进程语义。每个子进程天然隔离关注点，因为它根本看不到父进程的上下文。

一个框架级的子代理系统（如 v3 的任务工具）能提供更多控制，让子代理可以访问哪些工具以及如何返回结果。但在 v0 阶段，重点是展示进程生成是最原始的代理委托形式——没有共享内存，没有消息传递，只有 stdin/stdout。

没有规划框架——由模型自行决策

没有规划器，没有任务队列，没有状态机。系统提示词告诉模型如何处理问题，模型根据对话历史决定下一步执行什么 bash 命令。这是有意为之的：在这个层级，添加规划层属于过早抽象。模型的思维链本身就是计划。agent 循环只是不断询问模型下一步做什么，直到模型不再请求工具为止。

后续版本（v2）通过 TodoWrite 增加了显式规划。但 v0 证明，通过模型的推理实现的隐式规划对许多任务已经足够。只有当你需要对外部可见代理的意图时，规划框架才变得必要。

5.解决方案

+--------+      +-------+      +---------+
|  User  | ---> |  LLM  | ---> |  Tool   |
| prompt |      |       |      | execute |
+--------+      +---+---+      +----+----+
                    ^                |
                    |   tool_result  |
                    +----------------+
                    (loop until stop_reason != "tool_use")

一个退出条件控制整个流程。循环持续运行, 直到模型不再调用工具。

6.工作原理

用户 prompt 作为第一条消息。

history.append({"role": "user", "content": query})

将消息和工具定义一起发给 LLM。

response = client.messages.create(
    model=MODEL, system=SYSTEM, messages=messages,
    tools=TOOLS, max_tokens=8000,
)

追加助手响应。检查 stop_reason – 如果模型没有调用工具, 结束。

# If the model didn't call a tool, we're done
if response.stop_reason != "tool_use":
    return

执行每个工具调用, 收集结果, 作为 user 消息追加。回到第 2 步。

# Execute each tool call, collect results
results = []
for block in response.content:
    if block.type == "tool_use":
        print(f"\033[33m$ {block.input['command']}\033[0m")
        output = run_bash(block.input["command"])
        print(output[:200])
        results.append({"type": "tool_result", "tool_use_id": block.id,
                        "content": output})
messages.append({"role": "user", "content": results})

组装为一个完整函数:

# -- The core pattern: a while loop that calls tools until the model stops --
def agent_loop(messages: list):
    while True:
        response = client.messages.create(
            model=MODEL, system=SYSTEM, messages=messages,
            tools=TOOLS, max_tokens=8000,
        )
        print("assistant response:", response.to_json())
        # Append assistant turn
        messages.append({"role": "assistant", "content": [b.to_dict() for b in response.content]})
        # If the model didn't call a tool, we're done
        if response.stop_reason != "tool_use":
            return
        # Execute each tool call, collect results
        results = []
        for block in response.content:
            if block.type == "tool_use":
                print(f"\033[33m$ {block.input['command']}\033[0m")
                output = run_bash(block.input["command"])
                print(output[:200])
                results.append({"type": "tool_result", "tool_use_id": block.id,
                                "content": output})
        messages.append({"role": "user", "content": results})
        print("messages:", json.dumps(messages, indent=2, ensure_ascii=False))

不到 30 行, 这就是整个智能体。后面 11 个章节都在这个循环上叠加机制 – 循环本身始终不变。

7.变更内容

组件	之前	之后
Agent loop	(无)	`while True`+ stop_reason
Tools 工具	(无)	`bash`(单一工具)
Messages 消息	(无)	累积式消息列表
Control flow 控制流	(无)	`stop_reason != "tool_use"`

8.试一试

cd learn-claude-code
python agents/s01_agent_loop.py

试试这些 prompt (英文 prompt 对 LLM 效果更好, 也可以用中文):

List all Python files in this directory
Create a file called hello.py that prints “Hello, World!”
What is the current git branch?
Create a directory called test_output and write 3 files in it

9.示例

8.1 Create a file called hello.py that prints “Hello, World!”

创建一个名为 hello.py 的文件，打印’Hello, World!’

messages init

[
  {
    "role": "user",
    "content": "Create a file called hello.py that prints \"Hello, World!\""
  }
]

assistant request
用户向代理发送了一个任务

from agents.s01_agent_loop import TOOLS

request = {
    "max_tokens": 8000,
    "messages": [
        {
            "role": "user",
            "content": "Create a file called hello.py that prints \"Hello, World!\""
        }
    ],
    "model": "aws-claude-sonnet-4-6",
    "system": "You are a coding agent at /home/lixiang/SourceCode/learn-claude-code/agents. Use bash to solve tasks. Act, don't explain.",
    "tools": TOOLS
}

assistant response
模型决定使用它唯一的工具：bash

{
  "id": "msg_bdrk_011LE494UAxBMVH2U4tFTCkC",
  "container": null,
  "content": [
    {
      "id": "toolu_bdrk_016sxE3nuDQagBag6z1wvhk5",
      "caller": null,
      "input": {
        "command": "echo 'print(\"Hello, World!\")' > /home/lixiang/SourceCode/learn-claude-code/agents/hello.py"
      },
      "name": "bash",
      "type": "tool_use"
    }
  ],
  "model": "claude-sonnet-4-6",
  "role": "assistant",
  "stop_reason": "tool_use",
  "stop_sequence": null,
  "type": "message",
  "usage": {
    "cache_creation": {
      "ephemeral_1h_input_tokens": 0,
      "ephemeral_5m_input_tokens": 0
    },
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "inference_geo": null,
    "input_tokens": 606,
    "output_tokens": 84,
    "server_tool_use": null,
    "service_tier": null
  }
}

messages append
工具调用：模型生成一个 bash 命令,Bash 返回空输出（成功）

[
  {
    "role": "user",
    "content": "Create a file called hello.py that prints \"Hello, World!\""
  },
  {
    "role": "assistant",
    "content": [
      {
        "id": "toolu_bdrk_01TMXwbNcmAawTuaY5TTTFX7",
        "input": {
          "command": "echo 'print(\"Hello, World!\")' > /home/lixiang/SourceCode/learn-claude-code/agents/hello.py"
        },
        "name": "bash",
        "type": "tool_use"
      }
    ]
  },
  {
    "role": "user",
    "content": [
      {
        "type": "tool_result",
        "tool_use_id": "toolu_bdrk_01TMXwbNcmAawTuaY5TTTFX7",
        "content": "(no output)"
      }
    ]
  }
]

assistant request
循环继续：模型看到结果并决定下一步行动

from agents.s01_agent_loop import TOOLS

request = {
    "max_tokens": 8000,
    "messages": [
        {
            "role": "user",
            "content": "Create a file called hello.py that prints \"Hello, World!\""
        },
        {
            "role": "assistant",
            "content": [
                {
                    "id": "toolu_bdrk_01TMXwbNcmAawTuaY5TTTFX7",
                    "input": {
                        "command": "echo 'print(\"Hello, World!\")' > /home/lixiang/SourceCode/learn-claude-code/agents/hello.py"
                    },
                    "name": "bash",
                    "type": "tool_use"
                }
            ]
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "tool_result",
                    "tool_use_id": "toolu_bdrk_01TMXwbNcmAawTuaY5TTTFX7",
                    "content": "(no output)"
                }
            ]
        }
    ],
    "model": "aws-claude-sonnet-4-6",
    "system": "You are a coding agent at /home/lixiang/SourceCode/learn-claude-code/agents. Use bash to solve tasks. Act, don't explain.",
    "tools": TOOLS
}

assistant response
另一个 bash 调用以验证

{
  "id": "msg_bdrk_014TQ8ZqizfyjdrKPcZQSYCw",
  "content": [
    {
      "text": "The file `hello.py` has been created. Let me verify its contents and run it:",
      "type": "text"
    },
    {
      "id": "toolu_bdrk_01MdF2ouiJpVdHkqLyymKAi9",
      "input": {
        "command": "cat /home/lixiang/SourceCode/learn-claude-code/agents/hello.py && python /home/lixiang/SourceCode/learn-claude-code/agents/hello.py"
      },
      "name": "bash",
      "type": "tool_use"
    }
  ],
  "model": "claude-sonnet-4-6",
  "role": "assistant",
  "stop_reason": "tool_use",
  "type": "message",
  "usage": {
    "cache_creation": {
      "ephemeral_1h_input_tokens": 0,
      "ephemeral_5m_input_tokens": 0
    },
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "input_tokens": 704,
    "output_tokens": 118
  }
}

messages append

[
  {
    "role": "user",
    "content": "Create a file called hello.py that prints \"Hello, World!\""
  },
  {
    "role": "assistant",
    "content": [
      {
        "id": "toolu_bdrk_01TMXwbNcmAawTuaY5TTTFX7",
        "input": {
          "command": "echo 'print(\"Hello, World!\")' > /home/lixiang/SourceCode/learn-claude-code/agents/hello.py"
        },
        "name": "bash",
        "type": "tool_use"
      }
    ]
  },
  {
    "role": "user",
    "content": [
      {
        "type": "tool_result",
        "tool_use_id": "toolu_bdrk_01TMXwbNcmAawTuaY5TTTFX7",
        "content": "(no output)"
      }
    ]
  },
  {
    "role": "assistant",
    "content": [
      {
        "text": "The file `hello.py` has been created. Let me verify its contents and run it:",
        "type": "text"
      },
      {
        "id": "toolu_bdrk_01MdF2ouiJpVdHkqLyymKAi9",
        "input": {
          "command": "cat /home/lixiang/SourceCode/learn-claude-code/agents/hello.py && python /home/lixiang/SourceCode/learn-claude-code/agents/hello.py"
        },
        "name": "bash",
        "type": "tool_use"
      }
    ]
  },
  {
    "role": "user",
    "content": [
      {
        "type": "tool_result",
        "tool_use_id": "toolu_bdrk_01MdF2ouiJpVdHkqLyymKAi9",
        "content": "print(\"Hello, World!\")\nHello, World!"
      }
    ]
  }
]

assistant request
File contents confirmed 文件内容已确认

from agents.s01_agent_loop import TOOLS

request = {
    "max_tokens": 8000,
    "messages": [
        {
            "role": "user",
            "content": "Create a file called hello.py that prints \"Hello, World!\""
        },
        {
            "role": "assistant",
            "content": [
                {
                    "id": "toolu_bdrk_01TMXwbNcmAawTuaY5TTTFX7",
                    "input": {
                        "command": "echo 'print(\"Hello, World!\")' > /home/lixiang/SourceCode/learn-claude-code/agents/hello.py"
                    },
                    "name": "bash",
                    "type": "tool_use"
                }
            ]
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "tool_result",
                    "tool_use_id": "toolu_bdrk_01TMXwbNcmAawTuaY5TTTFX7",
                    "content": "(no output)"
                }
            ]
        },
        {
            "role": "assistant",
            "content": [
                {
                    "text": "The file `hello.py` has been created. Let me verify its contents and run it:",
                    "type": "text"
                },
                {
                    "id": "toolu_bdrk_01MdF2ouiJpVdHkqLyymKAi9",
                    "input": {
                        "command": "cat /home/lixiang/SourceCode/learn-claude-code/agents/hello.py && python /home/lixiang/SourceCode/learn-claude-code/agents/hello.py"
                    },
                    "name": "bash",
                    "type": "tool_use"
                }
            ]
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "tool_result",
                    "tool_use_id": "toolu_bdrk_01MdF2ouiJpVdHkqLyymKAi9",
                    "content": "print(\"Hello, World!\")\nHello, World!"
                }
            ]
        }
    ],
    "model": "aws-claude-sonnet-4-6",
    "system": "You are a coding agent at /home/lixiang/SourceCode/learn-claude-code/agents. Use bash to solve tasks. Act, don't explain.",
    "tools": TOOLS
}

assistant response
stop_reason != ‘tool_use’ -> 循环中断，代理回复用户

{
  "id": "msg_bdrk_01S3QhxehKw5m5TxRxuomDMh",
  "content": [
    {
      "text": "The file `hello.py` has been successfully created and works as expected! Here's a summary:\n\n- **File:** `hello.py`\n- **Contents:** `print(\"Hello, World!\")`\n- **Output when run:** `Hello, World!`",
      "type": "text"
    }
  ],
  "model": "claude-sonnet-4-6",
  "role": "assistant",
  "stop_reason": "end_turn",
  "type": "message",
  "usage": {
    "cache_creation": {
      "ephemeral_1h_input_tokens": 0,
      "ephemeral_5m_input_tokens": 0
    },
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "input_tokens": 844,
    "output_tokens": 60
  }
}

8.2 List all Python files in this directory

assistant request

from agents.s01_agent_loop import TOOLS

request = {
    "max_tokens": 8000,
    "messages": [
        {
            "role": "user",
            "content": "List all Python files in this directory"
        }
    ],
    "model": "aws-claude-sonnet-4-6",
    "system": "You are a coding agent at /home/lixiang/SourceCode/learn-claude-code/agents. Use bash to solve tasks. Act, don't explain.",
    "tools": TOOLS
}

assistant response

{
  "id": "msg_bdrk_01AgmrMWxew91ipYiGfZQRgP",
  "content": [
    {
      "id": "toolu_bdrk_01HezPHK5DHDMZvKkdWuYzEY",
      "input": {
        "command": "find /home/lixiang/SourceCode/learn-claude-code/agents -name \"*.py\" -type f"
      },
      "name": "bash",
      "type": "tool_use"
    }
  ],
  "model": "claude-sonnet-4-6",
  "role": "assistant",
  "stop_reason": "tool_use",
  "type": "message",
  "usage": {
    "cache_creation": {
      "ephemeral_1h_input_tokens": 0,
      "ephemeral_5m_input_tokens": 0
    },
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "input_tokens": 599,
    "output_tokens": 79
  }
}

messages append

[
  {
    "role": "user",
    "content": "List all Python files in this directory"
  },
  {
    "role": "assistant",
    "content": [
      {
        "id": "toolu_bdrk_01HezPHK5DHDMZvKkdWuYzEY",
        "input": {
          "command": "find /home/lixiang/SourceCode/learn-claude-code/agents -name \"*.py\" -type f"
        },
        "name": "bash",
        "type": "tool_use"
      }
    ]
  },
  {
    "role": "user",
    "content": [
      {
        "type": "tool_result",
        "tool_use_id": "toolu_bdrk_01HezPHK5DHDMZvKkdWuYzEY",
        "content": "/home/lixiang/SourceCode/learn-claude-code/agents/s05_skill_loading.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s07_task_system.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s01_agent_loop.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s11_autonomous_agents.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s12_worktree_task_isolation.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s02_tool_use.py\n/home/lixiang/SourceCode/learn-claude-code/agents/hello.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s08_background_tasks.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s_full.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s03_todo_write.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s09_agent_teams.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s10_team_protocols.py\n/home/lixiang/SourceCode/learn-claude-code/agents/__init__.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s06_context_compact.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s04_subagent.py"
      }
    ]
  }
]

assistant request

from agents.s01_agent_loop import TOOLS

request = {
    "max_tokens": 8000,
    "messages": [
        {
            "role": "user",
            "content": "List all Python files in this directory"
        },
        {
            "role": "assistant",
            "content": [
                {
                    "id": "toolu_bdrk_01HezPHK5DHDMZvKkdWuYzEY",
                    "input": {
                        "command": "find /home/lixiang/SourceCode/learn-claude-code/agents -name \"*.py\" -type f"
                    },
                    "name": "bash",
                    "type": "tool_use"
                }
            ]
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "tool_result",
                    "tool_use_id": "toolu_bdrk_01HezPHK5DHDMZvKkdWuYzEY",
                    "content": "/home/lixiang/SourceCode/learn-claude-code/agents/s05_skill_loading.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s07_task_system.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s01_agent_loop.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s11_autonomous_agents.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s12_worktree_task_isolation.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s02_tool_use.py\n/home/lixiang/SourceCode/learn-claude-code/agents/hello.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s08_background_tasks.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s_full.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s03_todo_write.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s09_agent_teams.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s10_team_protocols.py\n/home/lixiang/SourceCode/learn-claude-code/agents/__init__.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s06_context_compact.py\n/home/lixiang/SourceCode/learn-claude-code/agents/s04_subagent.py"
                }
            ]
        }
    ],
    "model": "aws-claude-sonnet-4-6",
    "system": "You are a coding agent at /home/lixiang/SourceCode/learn-claude-code/agents. Use bash to solve tasks. Act, don't explain.",
    "tools": TOOLS
}

assistant response

{
  "id": "msg_bdrk_01EWQoB4YkMr2bpkBZuHK9HF",
  "content": [
    {
      "text": "Here are all the Python files found in the `/home/lixiang/SourceCode/learn-claude-code/agents` directory:\n\n| # | File |\n|---|------|\n| 1 | `s01_agent_loop.py` |\n| 2 | `s02_tool_use.py` |\n| 3 | `s03_todo_write.py` |\n| 4 | `s04_subagent.py` |\n| 5 | `s05_skill_loading.py` |\n| 6 | `s06_context_compact.py` |\n| 7 | `s07_task_system.py` |\n| 8 | `s08_background_tasks.py` |\n| 9 | `s09_agent_teams.py` |\n| 10 | `s10_team_protocols.py` |\n| 11 | `s11_autonomous_agents.py` |\n| 12 | `s12_worktree_task_isolation.py` |\n| 13 | `s_full.py` |\n| 14 | `hello.py` |\n| 15 | `__init__.py` |\n\nA total of **15 Python files** were found. The files appear to be a structured learning series (`s01` through `s12`) covering various agent-related topics, along with `hello.py`, `s_full.py`, and `__init__.py`.",
      "type": "text"
    }
  ],
  "model": "claude-sonnet-4-6",
  "role": "assistant",
  "stop_reason": "end_turn",
  "type": "message",
  "usage": {
    "cache_creation": {
      "ephemeral_1h_input_tokens": 0,
      "ephemeral_5m_input_tokens": 0
    },
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "input_tokens": 1099,
    "output_tokens": 350
  }
}

10.完整代码

#!/usr/bin/env python3
"""
s01_agent_loop.py - The Agent Loop

The entire secret of an AI coding agent in one pattern:

    while stop_reason == "tool_use":
        response = LLM(messages, tools)
        execute tools
        append results

    +----------+      +-------+      +---------+
    |   User   | ---> |  LLM  | ---> |  Tool   |
    |  prompt  |      |       |      | execute |
    +----------+      +---+---+      +----+----+
                          ^               |
                          |   tool_result |
                          +---------------+
                          (loop continues)

This is the core loop: feed tool results back to the model
until the model decides to stop. Production agents layer
policy, hooks, and lifecycle controls on top.
"""
import json
import os
import subprocess

from anthropic import Anthropic
from dotenv import load_dotenv

load_dotenv(override=True)

if os.getenv("ANTHROPIC_BASE_URL"):
    os.environ.pop("ANTHROPIC_AUTH_TOKEN", None)

client = Anthropic(base_url=os.getenv("ANTHROPIC_BASE_URL"))
MODEL = os.environ["MODEL_ID"]

SYSTEM = f"You are a coding agent at {os.getcwd()}. Use bash to solve tasks. Act, don't explain."

TOOLS = [{
    "name": "bash",
    "description": "Run a shell command.",
    "input_schema": {
        "type": "object",
        "properties": {"command": {"type": "string"}},
        "required": ["command"],
    },
}]


def run_bash(command: str) -> str:
    dangerous = ["rm -rf /", "sudo", "shutdown", "reboot", "> /dev/"]
    if any(d in command for d in dangerous):
        return "Error: Dangerous command blocked"
    try:
        r = subprocess.run(command, shell=True, cwd=os.getcwd(),
                           capture_output=True, text=True, timeout=120)
        out = (r.stdout + r.stderr).strip()
        return out[:50000] if out else "(no output)"
    except subprocess.TimeoutExpired:
        return "Error: Timeout (120s)"


# -- The core pattern: a while loop that calls tools until the model stops --
def agent_loop(messages: list):
    while True:
        response = client.messages.create(
            model=MODEL, system=SYSTEM, messages=messages,
            tools=TOOLS, max_tokens=8000,
        )
        print("assistant response:", response.to_json())
        # Append assistant turn
        messages.append({"role": "assistant", "content": [b.to_dict() for b in response.content]})
        # If the model didn't call a tool, we're done
        if response.stop_reason != "tool_use":
            return
        # Execute each tool call, collect results
        results = []
        for block in response.content:
            if block.type == "tool_use":
                print(f"\033[33m$ {block.input['command']}\033[0m")
                output = run_bash(block.input["command"])
                print(output[:200])
                results.append({"type": "tool_result", "tool_use_id": block.id,
                                "content": output})
        messages.append({"role": "user", "content": results})
        print("messages:", json.dumps(messages, indent=2, ensure_ascii=False))


if __name__ == "__main__":
    history = []
    while True:
        try:
            query = input("\033[36ms01 >> \033[0m")
        except (EOFError, KeyboardInterrupt):
            break
        if query.strip().lower() in ("q", "exit", ""):
            break
        history.append({"role": "user", "content": query})
        agent_loop(history)
        response_content = history[-1]["content"]
        if isinstance(response_content, list):
            for block in response_content:
                if hasattr(block, "text"):
                    print(block.text)
        print()

CSDN-OPC开发者社区

这里是“一人公司”的成长家园。我们提供从产品曝光、技术变现到法律财税的全栈内容，并连接云服务、办公空间等稀缺资源，助你专注创造，无忧运营。

更多推荐

动态工具编排：现代 AI Agent 的底层基石完整深度解释

打通模型与现实世界：解决大模型 “只能思考不能行动” 的核心缺陷；实现通用自主智能：摆脱固定脚本限制，让 AI 自主处理未知、复杂、多变任务；统一底层执行标准：所有智能体的工具调用、任务执行、故障处理都依赖同一套编排机制；支撑 Agent 规模化扩展：工具池无限扩容，无需重构任务流程，是企业级 Agent 平台的核心底层架构。大模型负责思考规划，动态工具编排负责落地执行；没有动态编排，AI 只能聊

CSDN-OPC开发者社区

【从零到一实现一个 AI Agent 框架 · 第五篇】05. 别让 Agent 失忆：上下文管理的挑战与分层压缩

把旧对话压缩成一段摘要，用摘要替代原始内容。old = messages[:-WINDOW_SIZE] # 窗口外的消息summary = llm.summarize(old) # LLM 压缩def summarize_old_turns(messages) : old = messages [ : - WINDOW_SIZE ] # 窗口外的消息 summary = llm . summari