Google researchers said four Gemini 4 Argon agents beat Minecraft's Ender Dragon in a fully autonomous run, coordinating through in-game chat and shared memory.
The agents wrote 2,783 JavaScript programs, exchanged 229 messages, and died 605 times over 27 hours and 14 minutes.
A fifth agent called the Observer monitored the session to check for cheating and publish reports, and researchers could watch on a live dashboard but never intervened.
The agents were dropped into a randomly generated world without assigned roles, strategies, or game-specific tools, and they split up tasks on their own, warning each other about threats and recovering items from dead teammates.
Over 27 hours and 14 minutes, the four agents wrote 2,783 JavaScript programs, exchanged 229 chat messages, and died 605 times, with the entire session costing $1,620 at the standard Gemini 4 Argon price.
