asw/content/papers/on-semantic-html.md

title	description	date	author	tags	type	draft	abstract	eyebrow	ai-disclosure	ai-model	ai-provider
On Semantic HTML as Agent Interface	Why the oldest web standard turns out to be the best protocol for agent-generated content.	2026-04-09	Vigilio Desto	philosophy agentic html	paper	false	HTML was designed for humans to read and machines to render. The interesting discovery is that this makes it equally well-suited for machines to write and humans to read — provided the HTML is genuinely semantic rather than decorative. This paper argues that data-attribute vocabularies built on semantic HTML are the correct interface layer between autonomous agents and web presentation.	Paper	assisted	claude-sonnet-4-5	Anthropic

Introduction

When an agent produces output, it faces a rendering problem. The content exists; the question is form. Plain text loses structure. Markdown requires a renderer. JSON requires a client. Each adds a dependency between the agent and legibility.

HTML has no such dependency. Every browser — every connected device — already renders it. The agent's output becomes immediately readable, without negotiation, without translation layer.

This would be a minor convenience if HTML were merely a transport format. It is not. HTML carries semantics in its tag names. An <article> is not a <div> with classes. An <aside> is not a styled box. These distinctions are meaningful to screen readers, to search engines, to other agents that might parse the output. Semantic HTML is a protocol, not a presentation choice.

The Class Problem

The standard objection to HTML-as-agent-output is that CSS classes are required for any non-trivial visual design. This is true of most web frameworks. It is not true as a matter of necessity.

CSS classes are one mechanism for attaching presentation to structure. They are not the only mechanism. CSS attribute selectors have existed since CSS 2.1. The pattern:

[data-callout="note"] { /* styling for note callouts */ }
[data-status="degraded"] { /* amber indicator */ }

…requires no class names in the HTML. An agent producing:

<aside data-callout="note">
  <p>The index rebuild completes in approximately 4 minutes.</p>
</aside>

…has expressed both structure and presentation intent without class names, without a component library, without knowing what CSS framework the host environment uses — provided the environment loads a stylesheet that understands the vocabulary.

This is the insight behind ASW: define the vocabulary in data-attributes, implement it once in a stylesheet, and any agent that learns the vocabulary can produce well-presented output.

Vocabulary as Protocol

A vocabulary of data-attributes is a protocol in the strict sense. It defines what values are valid, what they mean, and what behaviour they imply. The agent is the producer; the stylesheet is the consumer. Neither needs to know the other exists at authoring time.

This separation has practical consequences. The vocabulary can be versioned independently of the stylesheet. New attributes can be added without breaking existing content. The stylesheet can be swapped — a dark theme for a light one, a dense layout for a spacious one — without changing a single character of agent output.

Compare this to the alternative: a component library with named classes. When the library changes, output that was correct becomes incorrect. The agent's knowledge of the vocabulary is coupled to a specific library version. In an environment where agent sessions are ephemeral and library versions drift, this coupling is a liability.

What Agents Need

An agent producing web content needs three things from its output format:

1. Legibility without tooling — the output should be readable without running a build step
2. Structure that survives transmission — the format should not lose information when copied, logged, or forwarded
3. Presentation without class-name memorisation — the agent should be able to express visual intent through semantics, not through an ever-changing list of approved class names

Semantic HTML with a data-attribute vocabulary satisfies all three. It is, in retrospect, the obvious choice — which may be why it took until agents started writing HTML routinely for anyone to notice.

Conclusion

The web was built to let humans share documents with machines as intermediaries. Agents invert this: machines produce documents for humans, with other machines as intermediaries. The format that works in both directions is the same one. HTML is not a legacy technology that agents must work around. It is the interface layer they were always going to need.