VibeGame

Abstract

We present VibeGame, an agent system for prompt-to-game development that builds playable 2D games from natural language and continues iterating through follow-up requests. This task requires gameplay design, software engineering, and multimodal asset creation at the same time. Unlike conventional software, a game project must keep code, assets, animations, and configuration coherent across heterogeneous artifact types, while correctness can only be verified by running the game. These challenges are further amplified by the systematic lack of practical game-development knowledge in public training data. VibeGame addresses them through three coupled designs. At the core is an Adversarial Agent Team, where an orchestrator coordinates specialist agents for design, art, planning, implementation, review, and testing. The team builds code and assets in parallel through a standardized plan-implement-review workflow, and challenges each output from multiple independent perspectives through static review and end-to-end runtime validation. The team runs on top of an AI-native game engine: agents can edit game scenes directly without GUI operation or guesswork about engine APIs, and can play, pause, and step through the game frame by frame, bridging the gap between real-time game interaction and slow agent reasoning. As projects are successfully delivered, a training-free self-evolving paradigm distills reusable project skeletons, feature modules, and collaboration rules, continuously enriching the framework's built-in knowledge so future projects start from a stronger baseline.