# When the Agent Is the User MCP, as far as I can tell, is exactly what it looks like: the API layer for the world of AI and agents. An open protocol that gives an agent a standard plug into tools and data, the USB-C port of this era. Enterprises are wiring it into their own data, into public data, into the whole suite of SaaS tools they already contract, and the agent reaches through all of it without ever loading a page. Which raises the uncomfortable question for my discipline: what becomes of the webapp, the surface we spent two decades optimizing? When the agent is the user, pixels stop being the interface. Tool schemas and capability descriptions become the information architecture, documentation becomes the storefront, and the [[Thesis - Current tools are lossy membranes|lossy membranes]] between tools finally start to thin, because context moves tool to tool without a human courier retyping it. This rhymes with [[UX - Progressive Disclosure and the All-Purpose Prompt|the blank box problem]] from the other direction: there, humans could not see what the machine could do; here, the machine needs the product to disclose itself legibly, in text, to a reader that never sees the screen. What remains for humans is oversight, and I think that is a promotion, not a demotion. The webapp drifts from cockpit to control tower: approvals, exceptions, audit trails, the veto, the place where a person watches what their agents did and decides whether to keep trusting them. Per [[Thesis - Agents are for understanding systems; Personas are for understanding people|the agents and personas split]], the systems work migrates to the agents, and whatever interface survives gets to be more human than it was: fewer forms, more judgment. The UX job does not disappear. It moves to the two surfaces that matter now: the schema the agent reads and the trust surface the human keeps.