You learned your hands by using them. An agent gets a paragraph.

This article explores the fundamental challenge of integrating AI agents with physical devices. Unlike humans, who learn to navigate the world through trial and error, AI agents lack a physical developmental process. The author argues that because agents cannot safely experiment with real-world hardware—such as locks or industrial machinery—they rely on 'manifests' to understand device capabilities. However, current standards often conflate three distinct types of information: technical capabilities, the consequences of actions, and situational context. The author suggests that the industry must move beyond simple function signatures to create a more nuanced semantic layer. By distinguishing between what a device can do, whether an action is reversible, and whether it should be triggered in a specific situation, developers can build safer, more reliable systems. The piece introduces DoSync, an open protocol designed to address these gaps in agent-device communication.