stasiandrExperiments with AI agents in game dev

UI for agents — a solved problem?

My hypothesis was that UI Toolkit suits agentic UI development better than uGUI. I tested it on windows from a popup to a drag-and-drop inventory.

So here's what I wanted to try. My initial hypothesis was that UI Toolkit is a better fit for agentic UI development than uGUI.

How I decided to test it: I took windows of different complexity, from a popup to a shop. A confirmation popup, a to-do list and a small in-game shop with tabs and item previews. I wanted to see what mistakes each option would make, where things would go wrong, where the difficulties would be — basically, any differences at all.

The conditions are the same for both: bare Unity 6, Claude Code, Opus 5.5 and unity-mcp, nothing else. The agent gets a spec and a mockup and builds the window on its own from a single prompt; I don't step in at all. Then I go through the result myself against a checklist the agent never saw.

Mockups: popup, to-do list, shop
Mockups: popup, to-do list, shop

And in the end, after three runs, it turned out that both options work, equally well and without errors, so I had nothing to compare. Both agents scored 100% everywhere.

Shop: mockup, UI Toolkit, uGUI
Shop: mockup, UI Toolkit, uGUI
R-16 Shop on UI ToolkitPlayOpen project
R-13 Shop on uGUIPlayOpen project

So I decided to make an even harder window — an inventory with a tile grid where you play this kind of Tetris: drag and drop, rotation, stacks, a two-handed weapon that takes the shield off, tooltips with item comparison and so on. Seemingly a genuinely hard task, especially on the first try.

Inventory mockup: a helmet being dragged onto an occupied slot
Inventory mockup: a helmet being dragged onto an occupied slot

And Claude actually handled it too. On both frameworks, in about 15 minutes each.

Inventory: mockup, UI Toolkit, uGUI
Inventory: mockup, UI Toolkit, uGUI
R-20 Inventory on UI ToolkitPlayOpen project
R-19 Inventory on uGUIPlayOpen project

How cleanly it did these tasks within each framework, I won't judge, because that's a value judgment. What isn't a value judgment is that it did all of it, and it works.

So, in my view, UI in game dev is solved, at the very least. I honestly find it hard to come up with a UI that's even harder. I guess you could add some client-server interaction, but that would be contrived, and for some reason I'm not sure Claude couldn't handle that too.

So UI looks like a solved problem. The next question is simply how to stay out of Claude's way in large existing projects.

That's it for this note. More conclusions to come in full articles.