Finally, a coding agent that ships complete browser games

What it does OpenGame generates complete, playable web games from a single text prompt. It focuses on the integration layer—cross-file state wiring, scene coherence, and game loops—where generalist agents usually collapse. The framework runs on GameCoder-27B, a 27B-parameter code model trained specifically for game engines via continual pre-training, supervised fine-tuning, and execution-grounded reinforcement learning.

The interesting bit The agent treats game building as a reusable skill. Its “Game Skill” accumulates project skeletons and a living protocol of verified fixes, so it scaffolds stable architectures instead of patching syntax in isolation. The authors also built OpenGame-Bench, which boots builds in a headless browser and scores interactive playability—Build Health, Visual Usability, Intent Alignment—using a vision-language model, because static linting won’t catch a broken boss fight.

Key highlights

• End-to-end browser games across genres: platformers, tower defense, card battlers, twin-stick shooters
• GameCoder-27B: a 27B model trained for game engine mastery via three-stage continual pre-training, SFT, and execution-grounded RL
• Reusable “Template Skill” library and “Debug Skill” protocol that track verified cross-file fixes across runs
• OpenGame-Bench evaluates via headless browser execution and VLM judging, not just static code checks
• Six fully playable demos with source archives, generated from single prompts

Verdict Grab it if you study agentic code generation or game prototyping. Skip if you need general app scaffolding or native engine output—this is explicitly web-only.