一、简介
Agent 代码沙箱是专为 AI Agent 场景设计的新一代运行时基础设施，提供安全隔离的云端执行环境，可防止 Agent 访问或篡改系统外资源。同时，支持多种编程语言，具备快速启动、高并发等特性，能满足多样化的 AI 任务需求。
llm-sandbox 是一个轻量级且便携的沙盒环境，专门设计用于在安全隔离的环境中运行大语言模型（LLM）生成的代码，该开源项目通过 Docker 容器提供了一个易于使用的接口，用于设置、管理和执行代码，简化了运行 LLM 生成代码的过程。
Langchain 是一个用于开发由大型语言模型（LLM）驱动的应用程序的框架，能简化 LLM 应用生命周期各阶段，可借助其开源组件、第三方集成构建有状态智能体。
二、环境准备
·已创建并部署好 TKE 集群，如果你还没有集群，请参见 创建集群。
·已创建节点池，并且节点池内有至少2个节点，机型推荐 SA5.LARGE8。
三、环境验证
验证 Python 版本，推荐版本3.11.6。
python –version
预期结果：
Python 3.11.6
验证 pip 版本，推荐版本23.3.1，预期结果：
pip 23.3.1 from /usr/lib/python3.11/site-packages/pip (python 3.11)
如果上述依赖不存在，可以运行以下命令安装依赖。
yum install python3 python3-pip -y #安装python3和pip
安装 llm-sandbox、langchain 和 langchain_openai。
pip install 'llm-sandbox[k8s]' langchain langchain_openai #安装依赖
四、构建一个工具调用的 Agent
我们将在集群中使用 Langchain 框架构建一个可以调用代码沙箱工具的 Agent，并且可以在沙箱中编写代码并返回结果。
五、步骤1：定义沙箱工具
代码沙箱提供执行代码的环境，并使用 K8s 作为后端，pod 作为执行代码的隔离环境。
@tooldef run_code(lang: str, code: str, libraries: list | None = None) -> str: """Run code in a sandboxed environment. :param lang: The language of the code, must be one of ['python', 'java', 'javascript', 'cpp', 'go', 'ruby']. :param code: The code to run. :param libraries: The libraries to use, it is optional. :return: The output of the code. """ with SandboxSession(lang=lang,backend=SandboxBackend.KUBERNETES,kube_namespace="default", verbose=False) as session: return session.run(code, libraries).stdout
六、步骤2：接入模型
场景接入使用 OpenAI 兼容接口的大模型，可以从 Deepseek 等模型服务提供商处获取 API_KEY，或在 TKE 上部署自建大模型。
llm = ChatOpenAI( model="xxx", # 填入模型名称，例如：deepseek-chat temperature=0, max_retries=2, api_key="xxxx", # 添入账户密钥，例如：sk-4***********************7 base_url="xxxxx",# 添入API端点地址，例如：https://api.deepseek.com)
七、步骤3：构建工具调用 Agent
使用 Langchain 构建工具调用 Agent，选择代码沙箱工具为可用的工具。
prompt = hub.pull("hwchase17/openai-functions-agent")tools = [run_code]agent = create_tool_calling_agent(llm, tools, prompt) # type: ignore[arg-type]agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True) # type: ignore[arg-type]
八、验证代码沙箱
设计如下四个执行代码的测试用例：
output = agent_executor.invoke({ "input": "Write python code to calculate Pi number by Monte Carlo method then run it." }) logger.info("Agent: %s", output) output = agent_executor.invoke({"input": "Write python code to calculate the factorial of a number then run it."}) logger.info("Agent: %s", output) output = agent_executor.invoke({"input": "Write python code to calculate the Fibonacci sequence then run it."}) logger.info("Agent: %s", output) output = agent_executor.invoke({"input": "Calculate the sum of the first 10000 numbers."}) logger.info("Agent: %s", output)
预期结果：
完成上述步骤执行，可以在集群中运行代码，查看运行效果，示例的效果如下：
> Entering new AgentExecutor chain…HTTP Request: POST http://xxxx/v1/chat/completions "HTTP/1.1 200 OK"Invoking: `run_code` with `{'lang': 'python', 'code': 'import random\\n\\ndef estimate_pi(n):\\n num_point_circle = 0\\n num_point_total = 0\\n \\n for _ in range(n):\\n x = random.uniform(0,1)\\n y = random.uniform(0,1)\\n distance = x**2 + y**2\\n if distance <= 1:\\n num_point_circle += 1\\n num_point_total += 1\\n \\n return 4 * num_point_circle/num_point_total\\n\\nprint(estimate_pi(1000000))'}`3.142396HTTP Request: POST http://xxxxx/v1/chat/completions "HTTP/1.1 200 OK"The calculated value of Pi using the Monte Carlo method with 1,000,000 sample points is approximately **3.142396**.This is a stochastic method, so the result will vary slightly each time you run it. The accuracy generally improves with a larger number of sample points.> Finished chain.Agent: {'input': 'Write python code to calculate Pi number by Monte Carlo method then run it.', 'output': 'The calculated value of Pi using the Monte Carlo method with 1,000,000 sample points is approximately **3.142396**.\\n\\nThis is a stochastic method, so the result will vary slightly each time you run it. The accuracy generally improves with a larger number of sample points.'}> Entering new AgentExecutor chain…HTTP Request: POST http://xxxxx/v1/chat/completions "HTTP/1.1 200 OK"Invoking: `run_code` with `{'lang': 'python', 'code': 'def factorial(n):\\n if n == 0:\\n return 1\\n else:\\n return n * factorial(n-1)\\n\\nnumber = 5\\nprint(f"The factorial of {number} is {factorial(number)}")'}`The factorial of 5 is 120HTTP Request: POST http://xxxxx/v1/chat/completions "HTTP/1.1 200 OK"Here is the Python code to calculate the factorial of a number:“`pythondef factorial(n): if n < 0: return "Factorial is not defined for negative numbers" elif n == 0 or n == 1: return 1 else: result = 1 for i in range(2, n + 1): result *= i return resultnumber = 5print(f"The factorial of {number} is {factorial(number)}")“`And as shown in the output, the factorial of 5 is 120.> Finished chain.Agent: {'input': 'Write python code to calculate the factorial of a number then run it.', 'output': 'Here is the Python code to calculate the factorial of a number:\\n\\n“`python\\ndef factorial(n):\\n if n < 0:\\n return "Factorial is not defined for negative numbers"\\n elif n == 0 or n == 1:\\n return 1\\n else:\\n result = 1\\n for i in range(2, n + 1):\\n result *= i\\n return result\\n\\nnumber = 5\\nprint(f"The factorial of {number} is {factorial(number)}")\\n“`\\n\\nAnd as shown in the output, the factorial of 5 is 120.'}> Entering new AgentExecutor chain…HTTP Request: POST http://xxxxx/v1/chat/completions "HTTP/1.1 200 OK"Invoking: `run_code` with `{'lang': 'python', 'code': '\\ndef fibonacci(n):\\n if n <= 0:\\n return []\\n elif n == 1:\\n return [0]\\n elif n == 2:\\n return [0, 1]\\n else:\\n fib_seq = [0, 1]\\n for i in range(2, n):\\n fib_seq.append(fib_seq[i-1] + fib_seq[i-2])\\n return fib_seq\\n\\n# Calculate the first 10 numbers in the Fibonacci sequence\\nresult = fibonacci(10)\\nprint(result)\\n'}`[0, 1, 1, 2, 3, 5, 8, 13, 21, 34]HTTP Request: POST http://xxxx/v1/chat/completions "HTTP/1.1 200 OK"Invoking: `run_code` with `{'lang': 'python', 'code': 'def fibonacci(n):\\n fib_sequence = [0, 1]\\n while len(fib_sequence) < n:\\n fib_sequence.append(fib_sequence[-1] + fib_sequence[-2])\\n return fib_sequence[:n]\\n\\nprint(fibonacci(10))'}`responded: The previous attempt to run the code did not produce any output. I will now write the Python code to calculate the Fibonacci sequence and execute it properly.[0, 1, 1, 2, 3, 5, 8, 13, 21, 34]HTTP Request: POST http://xxxx/v1/chat/completions "HTTP/1.1 200 OK"Here is the Python code to calculate the Fibonacci sequence up to the 10th term:“`pythondef fibonacci(n): fib_sequence = [0, 1] while len(fib_sequence) < n: next_value = fib_sequence[-1] + fib_sequence[-2] fib_sequence.append(next_value) return fib_sequence[:n]# Calculate the first 10 Fibonacci numbersresult = fibonacci(10)print(result)“`When executed, this code produces the following output:[0, 1, 1, 2, 3, 5, 8, 13, 21, 34]> Finished chain.Agent: {'input': 'Write python code to calculate the Fibonacci sequence then run it.', 'output': 'Here is the Python code to calculate the Fibonacci sequence up to the 10th term:\\n\\n“`python\\ndef fibonacci(n):\\n fib_sequence = [0, 1]\\n while len(fib_sequence) < n:\\n next_value = fib_sequence[-1] + fib_sequence[-2]\\n fib_sequence.append(next_value)\\n return fib_sequence[:n]\\n\\n# Calculate the first 10 Fibonacci numbers\\nresult = fibonacci(10)\\nprint(result)\\n“`\\n\\nWhen executed, this code produces the following output:\\n[0, 1, 1, 2, 3, 5, 8, 13, 21, 34]'}> Entering new AgentExecutor chain…HTTP Request: POST http://xxxx/v1/chat/completions "HTTP/1.1 200 OK"Invoking: `run_code` with `{'lang': 'python', 'code': 'print(sum(range(1, 10001)))'}`50005000HTTP Request: POST http://xxxxx/v1/chat/completions "HTTP/1.1 200 OK"The sum of the first 10,000 numbers is 50,005,000.> Finished chain.Agent: {'input': 'Calculate the sum of the first 10000 numbers.', 'output': 'The sum of the first 10,000 numbers is 50,005,000.'}
可参考的完整代码如下：
import loggingfrom langchain import hubfrom langchain.agents import AgentExecutor, create_tool_calling_agentfrom langchain_core.tools import toolfrom langchain_openai import ChatOpenAIfrom llm_sandbox import SandboxSession, SandboxBackendlogging.basicConfig(level=logging.INFO, format="%(message)s")logger = logging.getLogger(__name__)@tooldef run_code(lang: str, code: str, libraries: list | None = None) -> str: """Run code in a sandboxed environment. :param lang: The language of the code, must be one of ['python', 'java', 'javascript', 'cpp', 'go', 'ruby']. :param code: The code to run. :param libraries: The libraries to use, it is optional. :return: The output of the code. """ with SandboxSession(lang=lang,backend=SandboxBackend.KUBERNETES,kube_namespace="default", verbose=False) as session: return session.run(code, libraries).stdoutif __name__ == "__main__": # llm = ChatOpenAI(model="gpt-4.1-nano", temperature=0) llm = ChatOpenAI( model="qwen3-coder", temperature=0, max_retries=2, api_key="1d4a46b238eabef25101b5ff1dc36a150", base_url="http://49.233.239.193:60000/v1",) prompt = hub.pull("hwchase17/openai-functions-agent") tools = [run_code] agent = create_tool_calling_agent(llm, tools, prompt) # type: ignore[arg-type] agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True) # type: ignore[arg-type] output = agent_executor.invoke({ "input": "Write python code to calculate Pi number by Monte Carlo method then run it." }) logger.info("Agent: %s", output) output = agent_executor.invoke({"input": "Write python code to calculate the factorial of a number then run it."}) logger.info("Agent: %s", output) output = agent_executor.invoke({"input": "Write python code to calculate the Fibonacci sequence then run it."}) logger.info("Agent: %s", output) output = agent_executor.invoke({"input": "Calculate the sum of the first 10000 numbers."}) logger.info("Agent: %s", output)
九、pip 安装时无法找到 llm-sandbox 包？
Python 和 pip 版本过低，更新 python 和 pip 的版本，推荐 python3.11 以上。
十、运行测试代码时，遇到权限报错返回？
开启集群内网访问，并更新 config 文件。