Researchers introduce GameXpert-Bench, an evaluation framework with three benchmark tracks assessing coding agents across the game-development lifecycle: initial creation (GameGen, 97 tasks), bug diagnosis and repair (GameFix, 100 tasks), and iterative optimization (GameOpt, 17 chains). Evaluation uses live game interaction, deterministic behavioral tests, and final-product criteria with regression checks across eleven game genres. The findings show current agents excel at producing playable foundations and implementing explicit requirements but struggle significantly with defect detection and maintaining functionality through successive modifications.
