How have AI agents been enabled to generate good UI code starting today?
AI agents were enabled to create better UI by using a multi-agent feedback loop called a 'hill climb.' Multiple LLMs (including Opus, Codex, and Soul) serve as a review council that evaluates visual diff screenshots, rates design quality out of 10, and directs code edits through dozens of iterative rounds until target design scores are achieved.
Evidence
Title — claimed “Agents Sucked At UI (Until Today)”
Basis — The video demonstrates an automated multi-agent workflow that sends UI screenshots to multiple LLM models for critique. Over 18 hours and dozens of review iterations, the workflow systematically improved font hierarchy, radius, spacing, and contrast, taking evaluated UI design scores from 6/10 to 9+/10.








