VibeGame

We present VibeGame, an agent system for Prompt-to-Game Development that builds playable 2D games from natural language and iteratively refines them through follow-up requests. This task simultaneously demands gameplay design, software engineering, and multimodal asset creation. Unlike conventional software, game projects require code, assets, animations, and configurations to maintain mutual consistency across heterogeneous artifact types, and correctness can only be verified by actually playing the running game, a real-time system that responds in milliseconds while an agent's reasoning is orders of magnitude slower. These challenges are compounded by the systematic absence of game development practice knowledge in public training corpora. VibeGame addresses these issues through three coupled designs. The system is built around an Adversarial Agent Team centered on an orchestrator that coordinates multiple specialized agents, separating generation from review and challenging every output through static review and end-to-end runtime verification from multiple independent perspectives. This team runs on top of an AI-native game engine that represents the whole project as structured text editable without GUI interaction, exposes frame-synchronous control so that agents can play-test the real-time game at their own pace, and opens its full source code for agents to read. As projects are successfully delivered, a training-free self-evolution mechanism distills reusable project skeletons, modules, and collaboration contracts as project priors, continuously enriching the framework's built-in knowledge so that subsequent projects start from a higher baseline. We demonstrate the system's end-to-end capabilities through multiple games of different types, covering creation of complete playable projects from text-only or multimodal input, as well as editing scenarios including IP transfer, genre transfer, and rule restructuring on existing projects.