In 2026 Devoxx Belgium ran a competition: build a game set at the conference, starting from the venue floor plans and the models of three robots. The brief treated AI not as a shortcut but as something to experiment with — a craft to master.

So I started building, and won the competition. Every line of code was written by an agent. Every decision was mine.

This is the short version of what that process was actually like.

First rule: find the game before you build it. Twelve throwaway prototypes, gameplay proved in 2D, then 2.5D and 3D versions explored and refined in parallel before committing to one.

Then the gauntlet loop, which did the heavy lifting. A lead agent splits a goal into pieces; each piece gets a builder and a critic who never sees the builder’s reasoning — only a screenshot, judged against a concrete reference. Not “make it look professional”, but a real game as the benchmark. The biggest gap goes back to the builder. Repeat.

It worked remarkably well for graphics. It did not work for fun. Puzzles, pacing, and the bugs that only appear when a human actually plays — those came back to me every time.

You get the loop, the planning, what to do when the token quota runs out, and the move that saved me most often: when the agent couldn’t get the floor plan right, the fix wasn’t a better prompt: it was asking it to build me an interactive map editor.

Plus the best bugs, and a live demo.

Short presentation at Devoxx 2026.
https://www.youtube.com/watch?v=NT0wyPM58G8

Short session, 20 min preferred.