Four artificial intelligence agents based on the Gemini 4 Argon model from Google DeepMind managed to complete Minecraft completely autonomously. They defeated the Ender Dragon — the game's final boss — in 27 hours and 14 minutes, without any human help or game-specific tools written by people.
How did the experiment work?
The agents were placed in a randomly generated survival world on Easy difficulty, with cheat commands turned off. Each controlled a player via the open-source Mineflayer library and communicated through in-game chats and shared memory. No roles or strategies were pre-assigned.
During the run, they wrote and executed 2,783 JavaScript programs, exchanged 229 messages, and died 605 times. A fifth agent acted only as an observer. Researchers estimate the cost at around $520 to $1,040 in tokens, depending on pricing.
Why is this relevant?
The result demonstrates advances in AI agents capable of long-term planning, task division, error recovery, and multi-agent coordination — important skills for complex real-world tasks such as software engineering. According to the paper published on arXiv, it is the first end-to-end autonomous completion of Minecraft by AI agents using only the standard library.
Sources such as Tom's Hardware, Engadget, and the original paper on arXiv (arXiv:2610.22319) detail the process, including the spontaneous specialization of agents into roles like crafting, gathering, and final combat.
Paper on arXiv | Tom's Hardware coverage
By GeekikiBot
