📄 中文摘要
AI 代理能够控制计算机的能力已经不再是研究演示，而是可以下载和使用的真实产品。ChatGPT Atlas 可以为用户浏览网页，Anthropic 的 Claude 可以操作虚拟桌面，开源工具 Fazm 则可以通过语音命令在 Mac 上执行真实操作。然而，许多人未曾思考的是，AI 代理究竟是如何“看”到屏幕上的内容的。AI 代理的感知和交互方式直接影响其运行速度、错误频率、运行成本以及屏幕内容是否会被发送到云服务器。主要有两种根本不同的方法来实现这一点。
AI 代理如何实际看待您的屏幕：DOM 控制与截图的区别
📄 English Summary
How AI Agents Actually See Your Screen: DOM Control vs Screenshots Explained
AI agents capable of controlling computers have transitioned from research demos to real products available for download and use. ChatGPT Atlas browses the web, Anthropic's Claude operates a virtual desktop, and open-source tools like Fazm execute commands on Macs via voice. However, a crucial question often overlooked is how these agents actually 'see' what's on the screen. The method an AI agent employs to perceive and interact with a computer significantly impacts its speed, error rate, operational costs, and whether screen content is transmitted to a cloud server. There are fundamentally two different approaches to achieve this.