Four Google Gemini 4 agents beat Minecraft with no human intervention

Google researchers have announced in a new paper that four agents running the company's Gemini 4 Argon AI model completed the video game Minecraft by defeating its final boss, the Ender Dragon, with no human intervention. The experiment suggests that general-purpose agentic AI models can now handle complex games that once required custom-built software.

How the experiment worked

The team placed four agents in a sandboxed, randomized Minecraft survival world. No roles, strategies or game-specific tools were assigned. Each agent controlled its own player through the open-source Mineflayer library and coordinated with teammates via in-game chat and shared memory.

The agents quickly agreed to divide up the work. They warned one another about threats, passed on lessons and tips, shared coordinates and even retrieved items from fallen teammates. Roles emerged over time: "janesick" made tools, "gyldenia" gathered food and supplies, "auratian" lit a key portal, and "bulicame" built shelters against Endermen and landed the final blow.

The numbers

The whole process took 27.23 hours, most of it spent on the Ender Dragon fight. According to the researchers, the agents wrote and ran 2,783 programs, exchanged 229 messages and died 605 times. A fifth agent, the "Observer," acted as an independent witness, checking for cheating and publishing live and summary reports. The agents consumed 1.35 billion input tokens, costing $810 at Gemini's list price or $1,620 at the standard price.

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