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.
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.