一句话：提出首个多智能体端到端框架 UniGen，从自然语言自动生成可运行的 3D 游戏工程，全程无需手写代码。
3D 游戏开发要求同时掌握编程与引擎配置，门槛高，非专业用户难以参与。现有自动化方法存在三点不足：只覆盖 2D 内容或孤立代码片段；生成的组件需人工集成进游戏引擎；交互逻辑与状态转移管理效果不佳。核心缺口在于如何把 MLLM 输出转化为 Unity、Unreal 等引擎中可直接运行的生产级工程。为此作者提出 UniGen：规划智能体将自然语言需求转为结构化蓝图，生成智能体产出可执行 C# 脚本，自动化智能体负责组件绑定与场景构建，调试智能体进行实时错误修正。在三个游戏原型上评估，UniGen 完全免除手工编码，平均开发延迟降低 91.4%，实现与提示模板已开源。
3D game development demands expertise across programming and engine configuration, limiting non-professional participation. Existing automated approaches face three challenges: (1) scope limited to 2D content or isolated code snippets; (2) manual integration of generated components into game engines; and (3) suboptimal interactive game logic and state-transition management. A critical gap remains in translating MLLM outputs into production-ready, executable projects within engines like Unity and Unreal Engine. We introduce UniGen, the first end-to-end multi-agent framework for automating runnable 3D game development from natural language. UniGen employs a Planning Agent to convert requirements into structured blueprints, a Generation Agent to produce executable C# scripts, an Automation Agent for component binding and scene construction, and a Debugging Agent for real-time error correction. Evaluated on three game prototypes, UniGen eliminates manual coding and reduces mean development latency by 91.4%. The agent implementation and prompt templates are available at https://github.com/yxwan123/UniGen.