Agent-generated UI passes the screenshot test and fails the tab key

I run a design studio, and for the past year most of the UI I review was

written by an agent. The failure pattern is consistent, and it is not the

one I expected.

The output looks fine. Spacing is even, the type scale holds together, it

screenshots well. Then you press Tab and the focus ring is gone. The submit

button is disabled and nothing tells you why. The only label is the

placeholder, which disappears the second you type into it. The error arrives

as a toast that leaves after four seconds.

None of that is a capability gap. Every one of those is a decision the agent

made from whatever it absorbed, with nothing to check itself against and

nothing telling it which of the two hundred things on that page actually

decides whether the user finishes the task.

Two questions I would like real answers to:

1. What is the one UX mistake your agent makes every single time, no matter

how you prompt it?

2. When you review or audit a page, which part would you hand to an agent

today, and which part would you never give up?

We are building Rubric for the first half of that problem. The pre-launch

page is up if you want to follow along, but I am here for the answers more

than the follows.