This paper presents an alternative approach to GUI agents by reformulating world modeling to predict action-induced future information in latent space, improving success rates and reducing inference latency.