A2UI Express: 3× Fewer Tokens for Generative UI (2026)

On October 27, 2026, Google announced A2UI Express, a new inference format for its Agent-to-User Interface protocol that reduces token consumption by 3x and latency by 3x compared to Direct JSON output. When paired with the Jev decision model, the reduction reaches 80x for complex generative UI workflows. For developers building AI-powered interfaces, this is a significant optimization. But for Markdown publishers, it signals something broader: the industry is converging on the idea that structured source content should be lightweight, and presentation should be derived, not duplicated.

This is the same principle that makes Markdown such a durable format for content creation. A Markdown file is the canonical source — small, human-readable, version-controllable. The formatted output you see on X Articles, LinkedIn, or Notion is derived from that source at publish time. A2UI Express applies this philosophy to generative UI: the AI agent produces a compact, schema-aware representation of the interface, and the rendering layer expands it into the full UI. The source stays lean. The output adapts to context.

What A2UI Express actually does

A2UI (Agent-to-User Interface) is Google's protocol for letting AI agents generate user interfaces dynamically. Instead of hard-coding every UI state, an AI agent describes what the interface should look like using a structured format, and the client renders it. The original Direct JSON approach worked, but it was verbose — every UI element, every property, every state transition required explicit JSON tokens.

A2UI Express replaces this with a compressed inference format. Instead of transmitting the full JSON structure for every element, Express transmits only the data differences needed to render each component. A button doesn't need its full schema description every time — it needs a type identifier and the properties that changed. A list doesn't need every item wrapped in boilerplate — it needs the data array and a reference to the list component template.

The result is a 3x reduction in tokens and a 3x reduction in latency. When combined with the Jev decision model — which pre-filters the component set to only what's needed — the reduction reaches 80x. For real-time AI interfaces, this is the difference between a laggy demo and a production-ready experience.

The parallel to Markdown publishing

The A2UI Express optimization pattern is structurally identical to the Markdown publishing workflow that md2rich enables. Consider the parallel:

In both cases, the source format is deliberately minimal. Markdown uses asterisks for bold, hashes for headings, and hyphens for lists — nothing more. A2UI Express uses compact type identifiers and delta encoding — nothing more. The complexity lives in the rendering layer, not the source. This separation of concerns is what makes both systems efficient, adaptable, and future-proof.

Why the source format still matters in an AI-generated world

There's a temptation to think that as AI agents get better at generating formatted output directly, the need for a human-readable source format like Markdown diminishes. The A2UI Express announcement actually argues the opposite. By optimizing the transmission format rather than eliminating the source format, Google is acknowledging that structured, canonical sources are essential — even when the output is machine-generated.

For Markdown publishers, this has practical implications. When you draft in Markdown, you are creating a source file that can be:

• Version-controlled in git, with meaningful diffs between drafts
• Collaborated on in Google Docs (which now natively edits .md files) or any plain-text editor
• Converted to rich text for X Articles, LinkedIn, Notion, Medium, Substack, and other platforms
• Repurposed into newsletters, slide decks, documentation, or social threads
• Audited for compliance, fact-checking, or editorial review without formatting noise

An AI agent that generates formatted output directly — skipping the Markdown source — produces content that is harder to review, harder to version, and harder to repurpose. The A2UI Express philosophy suggests that even in a world of generative UI, the source format should be structured, minimal, and human-readable. Markdown already is.

The efficiency connection: tokens, bandwidth, and attention

A2UI Express's 3x token reduction is impressive, but the underlying insight is more fundamental: don't transmit what the client already knows. A UI component's schema doesn't need to be sent every time — the client has it. Only the data changes matter.

Markdown applies the same insight to content publishing. A heading doesn't need font size, weight, color, and spacing attributes embedded in the source — the rendering platform already knows how to style a heading. Only the structural intent matters: "this is a heading." The platform handles the presentation. This is why a Markdown file can be a few kilobytes while the same content as a Word document or Google Doc might be megabytes — and why converting Markdown to rich text is nearly instantaneous.

For publishers who care about efficiency — whether that's token efficiency in AI workflows or bandwidth efficiency in content distribution — the lesson is the same: keep the source lean, let the renderer do the work.

What this means for md2rich users

If you're already using md2rich to convert Markdown to rich text for publishing, the A2UI Express announcement validates your workflow. You're already practicing the source-lean, renderer-rich pattern that Google is now applying to generative UI. Your Markdown drafts are the canonical source. md2rich is the client-side rendering layer that produces platform-ready rich text without a server round-trip.

The tools are different — A2UI Express is for AI agents generating UI, md2rich is for humans publishing content — but the philosophy is identical. And as AI agents increasingly assist with content creation, having a clean, structured Markdown source becomes even more valuable. It's the format that both humans and machines can read, edit, and convert.

For a practical example of how Markdown-to-rich-text conversion works in a real publishing workflow, see our guide on Markdown Everywhere 2026. And if you're exploring how AI agents and Markdown interact, our article on Markdown in SQLite for AI agents covers another angle on token-efficient structured content.

FAQ