Google DeepMind Reports Four Gemini Agents Autonomously Beat Minecraft

Google DeepMind researchers recently published results showing that four AI agents based on the Gemini 4 Argon model successfully defeated the Ender Dragon in the sandbox game Minecraft without real-time human intervention. This marks the first time an AI has achieved end-to-end autonomous completion under conditions with no pre-scouted maps and no manually written custom tools.

The experiment lasted 27 hours and 14 minutes, during which the AI agents spent approximately $520 in compute tokens, relying entirely on the open-source Mineflayer library to write and execute 2,783 JavaScript code snippets to control character actions. Researchers observed via dashboard only, without interference.

The AI exhibited "emergent behavior," spontaneously evolving division of labor, with some agents crafting tools, others gathering resources, and others building defensive structures. Despite the final victory, the team died 605 times, and 22% of the executed code failed. The experiment demonstrates potential for AI collaboration in complex multi-step tasks, offering reference value for real-world software engineering automation.