📄 中文摘要
在人工智能项目中，构建一个神代理（god-agent）往往会导致失败。神代理是指一个大型语言模型（LLM）承担多项职责，如路由、验证、工具调用、合成、格式化和错误恢复等，然而缺乏监督。这种设计虽然在演示中表现良好，但在实际应用中却容易导致系统崩溃。作者观察到，多个生产系统因相同的模式而失败，强调了神代理的不可扩展性和崩溃风险。为了解决这一问题，建议采用更为分散和可管理的代理模式，以提高系统的稳定性和可维护性。
监督者模式：为何神代理总是崩溃（以及该如何构建替代方案）
📄 English Summary
The Supervisor Pattern: Why God-Agents Always Collapse (and What to Build Instead)
Building a god-agent in AI projects often leads to failure. A god-agent refers to a large language model (LLM) that handles multiple responsibilities, such as routing, validation, tool calling, synthesis, formatting, and error recovery, without supervision. While this design may perform well in demos, it is prone to catastrophic failures in real-world applications. The author has observed multiple production systems fail due to the same pattern, highlighting the scalability issues and collapse risks associated with god-agents. To address this problem, a more decentralized and manageable agent model is recommended to enhance system stability and maintainability.