OpenGame: Open Agentic Coding for Games (arxiv.org)

OpenGame has been introduced as the first open-source agentic framework designed specifically for end-to-end web game creation. It addresses the limitations of existing Large Language Models (LLMs) and code agents that struggle with producing fully playable games from high-level designs, often resulting in issues like cross-file inconsistencies and logical errors. Centered around a unique capability called Game Skill, which includes a Template Skill for evolving project skeletons and a Debug Skill for maintaining verified fixes, OpenGame enables agents to construct stable architectures and systematically address integration errors rather than merely fixing syntax issues. Significantly, this framework leverages GameCoder-27B, a specialized code LLM that employs a sophisticated three-stage training regimen of pre-training, fine-tuning, and execution-grounded reinforcement learning. To evaluate the performance of generated games, OpenGame introduces OpenGame-Bench, which assesses game creation through metrics like Build Health and Visual Usability using headless browser execution and visual language model judging. Demonstrating its capabilities, OpenGame achieved a new state-of-the-art across 150 diverse game prompts, indicating a significant advancement in the potential of code agents to tackle complex, interactive applications, paving the way for more sophisticated AI-driven game development. The framework is set to be fully open-sourced, promising greater accessibility and collaborative potential within the AI/ML community.