Google Gemini 4 agents just beat Minecraft with no human intervention

Agents split up tasks, warned about threats and even recovered items from dead teammates.

Google

Google's Deepmind "AlphaGo" AI is famous for winning at Go, but the company's latest Gemini AI just mastered a far more complex game. Four Gemini 4 Argon agents beat Minecraft's final "Ender Dragon" foe with no human intervention by working individually while sharing tips, threats and mistakes, Google researchers revealed in a new paper. The experiment shows how general agentic AI models can now beat complex games, a feat that used to required custom AI software.

To tackle Minecraft, the Gemini 4 team set four AI agents into a sandboxed, randomized Minecraft survival world with a goal to defeat the Ender Dragon. No roles, strategies or game-specific tools were assigned. Each agent controlled a player using the open-source Mineflayer library and coordinated with teammates through in-game chat and shared memory. 

During gameplay, agents quickly elected to split up tasks. They warned each about threats, gave lessons and tips, provided coordinates and even recovered items from dead teammates. Each settled into roles: for example, "janesick" made tools, "gyldenia" gathered food and supplies, "auratian" lit a key portal and "bulicame" built Enderman shelters and landed the final blow. 

All told, the process took 27.23 hours, with the majority of time spent on the Ender Dragon. "The four agents wrote and ran 2,783 programs, exchanged 229 messages and died 605 times along the way," the team said. A fifth agent called "the Observer" monitored the session as an independent witness, both to check for cheating and to publish live and summary reports. The agents used 1.35 billion input tokens and the final cost was $810 at Gemini's list price or $1,620 at the standard price. 

"To our knowledge, this is the first end-to-end autonomous completion of Minecraft by AI agents, achieved using only the standard Mineflayer library without pre-scouted seeds or other game-specific tools written by humans," the team wrote. "Winning required the agents to keep a coherent plan across thousands of dependent steps for more than a day."