The Agentic UX Shift: Should AI Speak in Markdown or HTML?
There's a question moving through design conversations right now, and it sounds technical at first. When an AI agent makes something for a person to read, what should that something be made of?
Markdown, the simple syntax that gave us bold and bullet points? Or HTML, the older language that built the web?
It started with a post by Andrej Karpathy and grew into an essay by an Anthropic engineer named Thariq Shihipar.[1] What looked like a small format debate opened into something bigger: what is AI output, really?
Markdown was made for writing. HTML was made for work.
That's the heart of it. When an agent gives you a thousand-line codebase review, or a dashboard of project status, or a comparison across many sources, the tools Markdown offers are thin. Bold text, headers, a few bullet points. HTML carries more. Sortable tables. Collapsible sections. Inline diffs. Interactive charts. Layouts that breathe.
Roughly thirty percent of the human cerebral cortex is given to visual processing.[2] Markdown barely knocks on that door.
Shihipar, who leads engineering on the Claude Code team, wrote an essay called The Unreasonable Effectiveness of HTML. He stopped using Markdown for AI output, and his reasoning is worth sitting with.[1] HTML output isn't where reading ends. It's where exploration begins.[2] A sortable, filterable table isn't a document. It's a tool. The reader doesn't just take the information in. They turn it over, sift it, follow what catches their eye.
The other side of the room
There are real arguments going the other way. Converting HTML into Markdown cuts token usage by about sixty-eight percent on clean content, and up to eighty-seven percent on messy web pages.[3] When the next reader of the output is another model in a chain, the lighter format wins.
There's also wisdom in constraint. Markdown can't be over-designed. HTML can, and often is. A small vocabulary protects you from a thousand small mistakes.[4]
What this means for designers
For us, the debate matters because it's really about what AI output is. If an agent speaks in Markdown, the output is text with formatting. If it speaks in HTML, the output is an interface.
That changes what we're responsible for.
In a Markdown world, the designer makes the container. The agent fills it with styled text inside the room you built. In an HTML world, the agent is making the room. Someone has to make sure that room is accessible, that it's navigable, that it carries the product's voice and pattern language with care. The output stops being content sitting inside your design. It becomes design, and it needs to be cared for as design.
This sits right next to the generative UI question from yesterday's briefing. They're the same question really, asked at different layers of the stack. Is what the agent makes a thing to read, or a place to act?
Where this is going
For now, the answer is probably both, depending on the moment and the user. But the way the wind is blowing, more agents will be making interfaces, and more interfaces will be made by agents. The work for designers is to meet that shift early. To build the systems, the components, the small standards that let agent output feel like a true part of the product rather than a guest passing through it.
The output is no longer the end of the conversation. It's the next room.
Sources[1] Let's Data Science — Anthropic Engineer Debates Markdown vs HTML for Agents
[2] Epsilla — The Agentic UX Shift: Why HTML is Replacing Markdown
[3] Beam.ai — HTML vs Markdown for AI Agents
[4] Cloudflare — Markdown for Agents