Gemini 4 Argon Agents Beat Minecraft Without Human Help

Researchers at Google DeepMind gave four agents running on the Gemini 4 Argon model a single task: beat Minecraft from the start all the way to killing the Ender Dragon. They had no predefined roles and got no in-game hints from humans. The agents wrote and debugged 2,783 programs on their own. They coordinated through in-game chat and a shared memory, and finished the game in 27 hours and 14 minutes. They died 605 times along the way, on easy difficulty. The authors call this the first fully autonomous Minecraft playthrough with no pre-set seeds and no special game tools written by humans.

Coordinating without a single predefined role is genuinely interesting engineering, and it is not really about the game. The agents split up tasks on their own and use shared memory to remember each other's mistakes. But 605 deaths and more than a day of effort to win a kids' game tell us nothing about whether this model is ready for real agentic work, where a mistake does not mean a respawn. It means broken production. In Claude Code, running on Opus and Fable, I solve tasks like this without that kind of barrage of attempts. Beating Opus and Fable on game benchmarks does not change that for Gemini, at least not yet.