I'd argue that content written by non-humans makes this even more important. You still likely have a need for a publishing / review workflow & attribution and a way to ensure standards are enforced and not just left to the end-user's own judgement.

Separating delivery from content management also reinforces this. Set up a governed content / user workflow and then let an agent focus on just the presentation after pulling in the content. I'd say this extends well beyond rest as well. A user interacting with something like a voice assistant in an agent might actually be using many interfaces at once (mcp from the assistant, HTML in a card, WebMCP alongside the HTML, browser open on another device for the human.) When that happens you want one canoninical system for the content.