📄 中文摘要
在生产环境中运行的 AI 代理可能会出现问题，例如陷入循环、错误调用工具或超出 API 预算。为了避免手动终止进程并希望没有出现故障，本文提供了一种有效的解决方案：构建一个反向代理，位于代理和大型语言模型提供者之间。所有请求都通过该代理流动，用户可以在 YAML 文件中定义策略。当触发策略时，请求会在到达模型之前被阻止。通过这种方式，代理可以在超出令牌预算或在特定时间窗口内发出过多请求时自动关闭，从而有效保护系统。
如何在 5 分钟内为你的 AI 代理添加一个杀开关
出处: How to Add a Kill Switch to Your AI Agent in 5 Minutes
发布: 2026年2月21日
📄 English Summary
How to Add a Kill Switch to Your AI Agent in 5 Minutes
In production, AI agents can encounter issues such as getting stuck in loops, hallucinating tool calls, or exceeding API budgets. To avoid the need for manual process termination and the hope that nothing broke, a practical solution is presented: building a reverse proxy that sits between the agent and the LLM provider. All requests flow through this proxy, where users can define policies in YAML. When a policy is triggered, the request is blocked before it reaches the model. This setup allows the agent to automatically shut down if it exceeds a token budget or makes too many requests within a specified time window, effectively safeguarding the system.