📄 中文摘要
AI 代理可能会出现不遵循指令的情况，导致其执行不该做的任务或忽视应做的任务。这种行为的根本原因并不在于模型或框架，而在于系统提示的设计。大型语言模型（LLM）本质上是填补空白的工具，能够根据角色和任务进行判断。然而，当用户设定的空白与模型的判断不一致时，就会出现问题。常见的空白模式包括能力与责任的不匹配，用户可能告知代理可以执行某项任务，但未明确何时不应执行该任务，从而导致代理的行为不符合预期。
为什么你的 AI 代理“失常”（问题不在模型）
出处: Why Your AI Agent "Misbehaves" (It's Not the Model)
发布: 2026年3月9日
📄 English Summary
Why Your AI Agent "Misbehaves" (It's Not the Model)
AI agents may stop following instructions, leading to them performing tasks they shouldn't or ignoring tasks they should. The root cause of this behavior lies not in the model or framework, but in the design of the system prompt. Large language models (LLMs) are inherently gap-fillers, capable of using judgment based on the role and task assigned. However, issues arise when the gaps set by the user conflict with the model's judgment. Common gap patterns include mismatches between capability and responsibility, where users inform the agent it can perform a task but fail to specify when it should refrain from doing so, resulting in unexpected behavior.