The AI plane is where agents have real access to infrastructure, and the controls around them are the whole point. Modding is where I try the same agents, and newer frontier models, on something I can afford to break. A mod that fails loses a save; a lab change that fails loses a weekend. So this is the lower-stakes bench: a place to experiment with agentic workflows, split work across subagents, put an independent reviewer over an unattended run, and find out what each model actually does with a game engine’s toolchain.

This page is the catalogue and the method: what the agents can touch, how a mod gets from a plan into the game, what has been built, and what had to be fixed on the way. The fixes are the useful part.