📄 中文摘要
在过去两年中，AI自动化、代理系统或智能工作流工具的开发者们普遍存在一个误解，即将大型语言模型（LLMs）视为完全功能的执行引擎。尽管LLMs能够编写代码、生成逐步工作流程、连接外部工具，并在几秒钟内返回“已完成任务”的响应，但这并不意味着它们可以替代传统的状态执行系统。尽管演示视频看起来令人印象深刻，早期测试似乎有效，但在真实的生产环境中，这种设置常常会面临幻觉、非确定性输出、状态管理失效等问题，导致一致性失败。
大型语言模型为何无法成为“执行实体”——一种基本范式的崩溃
📄 English Summary
Why LLMs Can Never Be "Execution Entities" — A Fundamental Paradigm Breakdown
In recent years, a widespread misconception has emerged among developers of AI automation, agent systems, and intelligent workflow tools: treating large language models (LLMs) as fully functional execution engines. While LLMs can write code, generate step-by-step workflows, connect to external tools, and return 'completed task' responses in seconds, this does not imply they can replace traditional stateful execution systems. Despite impressive demo videos and seemingly successful early tests, deploying this setup in real production environments often leads to consistent failures, including hallucinations, non-deterministic outputs, and broken state management.