GUI Agents for Continual Game Generation
The work investigates using GUI agents to improve interactive code generation by acting as both objective evaluators and subjective playtesters. It proposes two roles: an objective evaluator via PlaytestArena and a subjective playtester via Play2Code. Play2Code links game generation and playing in a loop, where a GUI agent observes gameplay to refine the next code iteration, achieving a 66.8% rubric pass-rate across experiments.