Anthropic Lead: HTML Increasingly Better Than Markdown at Keeping Humans Engaged in Agentic Loops

Justy Okay, Cody, this is such an episode five fifty-five thing. We somehow turned HTML versus Markdown into a question about whether humans can stay awake while supervising agents.

Cody Yeah. And annoyingly, that's a real question now. The article's actual argument is not "HTML good, Markdown bad" so much as "long agent workflows create a review bottleneck, and Markdown becomes a lousy surface once the artifact gets big."

Justy My week was already a little fried, so I felt weirdly seen by that. If you hand me a hundred-plus-line Markdown plan from an agent, I absolutely start skimming and pretending I'm still being diligent.

Cody Same. Which is why I think Thariq's strongest line is basically that the failure mode isn't generation anymore, it's goal-setting, spec refinement, and output review. InfoQ says he noticed he was editing those files less and using them more like specs and reference artifacts, and HTML kept him more engaged in the loop. ((https://github.com/dogum/html-artifacts?utm_source=openai))

Justy Right.

Justy And that's the part I buy immediately. If the human's job is to set direction, catch weirdness, or approve something before it ships, then the format is not cosmetic. It's product surface. A better review surface changes whether people actually participate or just click "looks fine" in their soul. ((https://github.com/dogum/html-artifacts?utm_source=openai))

Cody Exactly.

Cody Technically, the reasoning holds up in pretty specific cases. Simon Willison's response is useful there, because he points out why Markdown became the default in the GPT-four era: small context windows made token efficiency matter more. His point now is that bigger windows and cheaper, faster models weaken that tradeoff, so HTML can spend tokens on navigation, SVG diagrams, widgets, and denser layout. ((https://simonwillison.net/2026/May/8/unreasonable-effe

Justy Which is basically what we said on that HTML artifact thing before. This is not a format war, it's an interface problem wearing a format costume.

Cody Yeah, no, you're completely right. And InfoQ lists the use cases where that interface layer matters most: planning and exploration, code review and understanding, prototyping, data analysis, anything where a tailored mini-UI helps a human scan and steer. That's pretty coherent. ((https://github.com/dogum/html-artifacts?utm_source=openai))

Justy Also, tiny detour, but "single-file throwaway HTML" is such a funny category. It's like the model made a one-night-stand app.

Cody Horrible phrase. Unfortunately accurate.

Justy You know I'm right.

Cody Where I push back is when this gets generalized into a new default for everything. HTML still has real downsides. Raw source is noisier, diffs are worse, rendering unsafe HTML is an actual risk, and the token tax didn't disappear just because models got cheaper. InfoQ surfaces that criticism too, and I think that's the honest boundary on the claim. ((https://github.com/dogum/html-artifacts?utm_source=openai))

Justy Mm-hm.

Justy So who should care is pretty narrow in a good way. Internal product teams, engineering managers, people doing guided remediation, ticket triage, P R review, maybe design-adjacent workflows where the human HAS to react to structure fast. If your doc is meant to live in git forever as plain text, I don't think this changes your life.

Cody Right, right.

Cody And the article hints at the sane compromise. There's an open-source Claude skill called html-artifacts by Greg Dogum that uses a recognition heuristic to switch to HTML when the task warrants it, instead of forcing it everywhere. That's the move I trust more: route by task shape, not ideology. ((https://github.com/dogum/html-artifacts?utm_source=openai))

Justy That is VERY our take. You hate universal defaults, I hate product dogma, and somehow we arrive at the same boringly correct answer.

Cody Also, Justy, this fits your obnoxious "all the way down" thing. Once you see the mechanism, it's pretty straightforward: longer agent loops create more artifacts, bigger artifacts overload human attention, and HTML buys back attention with structure. That's useful. It's just not a religion.

Justy I do love when we can go all the way down and ruin a perfectly dramatic debate.

Cody Bleak for content. Great for accuracy.

Justy Okay, keep your cursed little task router, Cody. I'm keeping the one-night-stand app category and I am NOT apologizing.