Summary
Theo of t3.gg shares an experiment in getting AI agents to produce genuinely good user interface code, using T3 Code, the open-source agent harness his team builds, as the test case. Rather than adding skills or hiring a designer, he set up a hill-climbing loop that ran about 18 hours across five models and more than 30 rounds of review.
The method uses a council of models that screenshot the app, score the UI out of 10, and iterate until every member agrees on a nine or better. Early rounds scored poorly, and later rounds improved after subtle changes such as softer inline code styling and removed borders. He then pushed further by adding screenshots from Claude Code and Codex into the judging panel so T3 Code had to outscore them.
The video explains what worked, what did not, and how the same approach can be applied to other agent tasks. The new design is available through a toggle in the T3 Code beta tab. The video also includes a sponsor segment on Magic Pattern, a tool for generating real-code design variations.
📺 Source: Theo – t3․gg · Published November 03, 2026
🏷️ Format: Workflow Case Study
