Just give me the markdown
A while ago I was helping a friend with her CV, which meant reading her entire portfolio site. I did not have the time, and honestly I did not want to. So I put an agent on it. It read everything and gave me back one markdown file. I loaded that file into my assistant, and from there I could work with her whole body of work without ever browsing the site as she designed it.
Her site is good. That is not the point. The point is that for my use case, her site was in the way. I did not need the layout, the navigation, the typography choices. I needed the content, structured, and nothing else.
I think that experience is a preview of a real shift: the web splitting into data and rendering, with the rendering moving to the user's side. A site stops being only a finished surface and becomes, additionally, a data endpoint that your own tools consume, something like an MCP connector per website. The same canonical content, and each person consumes it in the form that serves them: your assistant reads it, your feed summarizes it, your screen renders it under your rules for how you want to consume information.
We have actually wanted this for a long time. RSS was exactly this instinct: just give me the content, skip the chrome. It worked, a generation of readers loved it, and it faded not because the desire went away but because the business of the web pushed the other way. What is new is that the rendering side got cheap. A UI can be generated in seconds and shaped to your preferences, which makes personal renderers viable in a way RSS readers never quite were. The signal-to-noise argument finally has the tooling it needed.
Now the tension, because I have heard where this argument goes when it runs unsupervised. When ChatGPT arrived, plenty of people concluded we would not need websites at all. No interfaces, just talk to your phone. That prediction aged badly, and it was always going to. Try shopping by voice. You want to see the products, compare them at a glance, click on things. We did not invent graphical interfaces by accident; direct manipulation exists because for many tasks it is simply the fastest channel between a human and a machine. The research community keeps reconfirming it: comparative studies find classic GUIs beat chat when the options are visible and the action is quick, while conversation wins for open-ended, multi-turn tasks. As one designer put it, a blank text box has no affordances; you should not have to prompt-engineer your way through life. Even the AI labs have quietly conceded the point, bolting visual interfaces back onto their chat products as fast as they can.
So the honest version of the thesis is not "the end of UI". It is that the use case defines the interaction. Sometimes that is a crafted, opinionated interface, and craft will keep winning there: shopping, exploration, anything where seeing and manipulating is the task. Sometimes it is a conversation. And sometimes, more often than the current web admits, it is neither: you just want the markdown. Research, comparison, feeding your own tools, catching up on content you trust. Those use cases are underserved by design, because every site assumes its surface is the only door.
The sites that win the next few years will have two doors: the crafted surface for humans who are browsing, and the data endpoint for the humans who sent their agents. Refusing the second door will not protect the first. It will just mean my agent reads your site anyway, badly, through the front window.