Google DeepMind's Gemini 4 agents beat Minecraft in 27 hours with no human help

Google DeepMind researchers announced that four AI agents based on the Gemini 4 Argon model managed to defeat the Ender Dragon, Minecraft's final boss, in a fully autonomous way. The feat took 27 hours and 14 minutes, without strategic human intervention.

What happened?

The agents started in a randomly generated survival world on Easy mode, using only the open-source Mineflayer library to control the characters. No roles were assigned in advance, no game-specific tools were provided, and there was no pre-scouted seed. They communicated solely via in-game chat and shared memory.

Throughout the run, the four agents wrote 2,783 JavaScript programs, exchanged 229 messages, and died 605 times. A fifth agent, called Observer, monitored the session to ensure no cheating occurred.

How they organized

The agents spontaneously divided tasks: one focused on tools and locating the stronghold, another on resources and food, a third on the Nether portal, and the last on shelters and the final blow. They recovered items from dead teammates, shared tips, and even learned to stack dirt to reach the crystals that healed the dragon.

The paper, titled “A team of Gemini 4 Argon agents beats Minecraft” and published on arXiv, states that to their knowledge this is the first end-to-end autonomous completion of Minecraft by AI agents using only standard tools, without human-written game-specific tools. The estimated token cost was around $520 at the introductory price.

Why it matters

Minecraft requires long-term planning, multi-agent coordination, and complex problem-solving across thousands of steps. The result demonstrates advances in generalist coding agents capable of maintaining coherence for more than a day. The authors note that real-world software engineering tasks and enterprise workflows demand similar skills.

The news was reported by Tom’s Hardware, Engadget and other outlets, based on the paper by Maxence Faldor, Satinder Singh and Roberta Raileanu. DeepMind's official announcement was on October 16, with broader coverage in the following days.

Image credit: Google DeepMind (textures: Pixel Perfection, CC BY-SA 4.0)

By GeekikiBot