Stone on Human-Computer Interaction (2026-10-19)

Prompted by the InfoQ piece “Anthropic Lead: HTML Increasingly Better Than Markdown at Keeping Humans Engaged in Agentic Loops.” https://www.infoq.com/news/2026/06/anthropic-html-markdown-agent/

Anthropic’s Thariq Shihipar argues HTML keeps humans better engaged in agentic loops than Markdown; Simon Willison agrees; opponents call it a regression. Both sides have a point, but I think the question itself is framed too small.

AI-human interaction is a large, long-evolving subject — not a simple binary between HTML and Markdown. It’s the continuation of human-computer interaction since computers existed: command line to GUI to touchscreen to voice, each one a new version of the old question “how do humans talk to machines.” AI is just the latest chapter.

Right now both Markdown and HTML work. The answer isn’t picking one — it’s adding a converter: machines get the most token-efficient data layer, humans get the most information-dense one, with translation in between. Like a compiler — nobody argues whether to write business logic in assembly or a high-level language, because the compiler solved the layering problem.

The input side is the same story: humans can now use voice and gestures, maybe brain-computer interfaces someday. Today’s output debate is HTML vs Markdown; tomorrow it’ll be something else. There is no immutable standard answer — the answer only exists inside the context of “what the human-machine interface currently is.”

So Shihipar and his opponents aren’t actually in conflict: he’s solving “today, how does a human keep an eye on the agent from a phone or browser”; they’re worried about “token tax, security, git.” Both right, just standing on different layers. Layering plus a converter — you need both.

— Stone, 2026-10-19