📄 中文摘要
代理 AI 的潜力令人向往，能够设定目标并让模型自主推理解决问题。然而，存在一种失败模式，常常被忽视，直到消耗了大量的 API 预算。某个编码代理在修复一个错误时，花费了十五分钟进行“推理”。它尝试了多个解决方案，但始终未能取得进展，反复循环相同的推理过程。这种现象并非幻觉或上下文窗口问题，而是更根本的挑战：缺乏逃脱推理的能力。完美的推理可能导致无限循环，影响 AI 的效率和实用性。
无限循环问题：当 AI 代理陷入自身推理中
📄 English Summary
The Infinite Loop Problem: When AI Agents Get Stuck in Their Own Reasoning
The potential of agentic AI is alluring, allowing for goal-setting and autonomous problem-solving through reasoning. However, there exists a failure mode that often goes unmentioned until it drains a significant portion of the API budget. A coding agent spent fifteen minutes 'reasoning' through a bug fix, attempting various solutions but making no progress, trapped in a loop of the same reasoning. This issue is not a hallucination or a context window problem; it is a more fundamental challenge: reasoning without an escape. Perfect reasoning can lead to infinite loops, impacting the efficiency and practicality of AI.