What they did
A Google DeepMind team, credited as Faldor et al., dropped four agents running on Gemini 4 Argon into a randomly generated Minecraft world. They gave them no roles, no strategy and no game-specific tools, and no human stepped in during the run. The agents acted by writing JavaScript programs to control the game, talked to each other in chat, split up tasks on their own, warned each other about threats, picked up items from teammates who had died and logged their mistakes in a shared memory. The harness is the team's own research setup.
How it went
The agents beat the game by slaying the Ender Dragon in one fully autonomous run lasting 27 hours and 14 minutes. Over that time they wrote 2,783 JavaScript programs, exchanged 229 chat messages and died 605 times. The report puts the cost of the whole session at $1,620 at the standard Gemini 4 Argon price due in 2027.
Worth knowing
This is a research result reported by The Decoder, not a hobbyist setup you can install. The deaths count shows how much trial and error a long autonomous run needs, and the cost figure uses a price that does not apply yet.