Using software without learning its interface

I use HubSpot every day, but I have no idea how to navigate HubSpot. I’m probably not going to learn.

Those statements should conflict. Increasingly, they don’t.

Phone numbers used to stick because I dialed them. Routes stuck because getting somewhere required paying attention. Contacts and GPS removed that repetition, and the knowledge stopped accumulating.

Agents may do the same thing to application interfaces.

I don’t want to open each tool just to use its agent. I want its data and capabilities available through the agents I’m already using.

For some users, an in-app agent will be helpful. For me, it is another stop. As Codex, Cowork, Copilot, and similar tools make other software easier to operate, I expect more people to use applications well without learning their interfaces.

The traditional UI probably doesn’t disappear. Its job changes. It becomes where we inspect the underlying record, resolve ambiguity, and verify ground truth when the agent’s answer needs checking.

That creates a difficult design problem. The UI can no longer be an application’s only useful surface.

That is why MCP belongs in product design, not at the bottom of the integration backlog.

An AI-first application needs to work when the user doesn’t open it and remain inspectable when they do.