- Jemima Abu said that, without guidance, a coding agent built its own versions of features that native HTML already provides, adding code and testing.
- Her fix is a mandatory accessibility skill favoring semantic HTML, though she said manual testing is still needed.
A coding agent given no accessibility guidance tends to build its own versions of features that browsers already provide. Jemima Abu made this argument in her talk on the AI Engineer show, drawing on a small experiment. In her account, the result is extra code and extra testing that a built-in HTML element would have avoided.
What was said
Even for a simple accordion (a panel that opens and closes) or a modal (a pop-up window), Abu said, the agent did not check whether a standard element already did the job. It wrote one from scratch. In her words, "it will reinvent the wheel as many times as it wants to".
The mechanism is simple. A native element arrives with its behavior built in. The show notes list semantics, keyboard interaction and expected browser behavior. A custom copy must rebuild each of those, and each rebuilt piece needs testing. Abu said this takes many more features and much more testing.
The show notes report that one run with an accessibility skill cost 58% less and took about a third of the time. Abu noted herself that the test was not controlled, so it does not show that accessible work is cheaper every time.
Her fix is a mandatory accessibility skill: standing instructions the agent follows when it writes interface code. It says to use semantic HTML, meaning markup that names what each part is, wherever possible. ARIA, an add-on set of accessibility labels, comes only when plain HTML cannot express the behavior. Anything a mouse can operate should work by keyboard, and the agent audits its code before running it.
Why it matters
Our reading: unguided agents add hidden cost by default. A demo looks the same whether the accordion is native or custom. The extra code, testing and upkeep show up later, and the team that owns the code pays for them.
For anyone who signs off agent-written interface code, her talk suggests four questions. Does the delivered accordion or modal use a built-in element, or a custom script? Can every control that works with a mouse also work with a keyboard? Is an accessibility rule part of the agent's standing instructions? Does an audit run before the code ships, with a person also testing by hand?
The other side
The evidence is thin. Abu called her experiment small, and the cost gap may owe something to how the two runs differed. She also said no single tool can make a site fully accessible, so manual testing is still needed.
Her talk was about agents as authors of code. She did not test whether agents browsing the web read semantic markup better. Linking her point to a machine-readable web is our extension, not something she showed.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: The Cost of Accessibility in AI — Jemima Abu, CAIS (2026-10-14)
