Google DeepMind 论文将 4 个 Gemini 4 Argon 智能体放入沙盒 Minecraft 生存世界,只给一个目标击败 Ender Dragon,未分配角色或预写游戏工具,智能体通过聊天和共享记忆自行组织分工。
New Google DeepMind paper shows 4 general-purpose coding agents beating Minecraft with no human help, writing every program and splitting the work themselves.
Finishing Minecraft means holding 1 plan across thousands of dependent steps, the same skill long software and business tasks need. Earlier Minecraft agents were trained on gameplay or handed game-specific planners.
Each Gemini 4 Argon agent got 1 goal, defeat the Ender Dragon, and a standard library that turns JavaScript into game actions. Nobody assigned roles or wrote game tools, so the team organized itself through chat and shared notes.
From a random world on Easy difficulty, they killed the dragon in 27 hours and 14 minutes, writing 2,783 programs and dying 605 times. It was a single run, so how often this setup succeeds is still unknown.
Before writing custom planners for a long task, try general coding agents with a plain API and a shared memory.
来源:Rohan Paul · x.com