📄 中文摘要
在使用 AI 编码代理如 Claude Code 或 Cursor 时，用户常常会遇到代理执行错误的情况，例如错误删除文件或破坏认证模块。用户通常会尝试回溯事件，查看对话记录和修改差异，试图找出问题所在。然而，当用户再次运行相同的操作时，代理却表现得截然不同。这种现象并非软件缺陷，而是 AI 系统固有的非确定性特征。非确定性使得 AI 代理在相同输入下可能产生不同的输出，给开发和调试带来了极大的挑战，影响了用户对 AI 系统可靠性的信任。
为什么无法重现 AI 代理的失败（这为何是个大问题）
📄 English Summary
Why You Can't Reproduce AI Agent Failures (And Why That's a Huge Problem)
Users of AI coding agents like Claude Code and Cursor often encounter situations where the agent performs incorrectly, such as deleting files it shouldn't or breaking authentication modules. Users typically attempt to trace back the events by reviewing conversations and diffs to understand what went wrong. However, when they try to rerun the same operations, the agent behaves completely differently. This phenomenon is not a bug but a fundamental characteristic of AI systems known as nondeterminism. Nondeterminism means that AI agents can produce varied outputs even with the same inputs, posing significant challenges for development and debugging, and undermining user trust in the reliability of AI systems.