Researchers from Google DeepMind placed four AI agents based on the Gemini 4 Argon model into a randomly generated Minecraft survival world and gave them a single goal: defeat the Ender Dragon, the game’s final boss.
What happened?
The agents controlled characters through the open-source Mineflayer library and coordinated exclusively via in-game chat and shared memory. No one assigned roles, strategies or game-specific tools. Every program was planned, written and debugged by the agents themselves.
In 27 hours and 14 minutes, the team wrote 2,783 programs, exchanged 229 messages and died 605 times — mostly to monsters. The Ender Dragon alone accounted for 69 deaths. According to the paper published on arXiv (arXiv:2610.22319), the authors believe 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 human-written game tools.
How it worked
The agents specialized on their own: one crafted tools and explored fortresses, another gathered supplies and destroyed crystals, another prepared the Nether portal and planned the crystals, and the last built Enderman shelters and delivered the final blow. They recovered items from fallen teammates and shared discoveries.
The estimated cost up to the win was about $520 at the introductory API price (or $1,040 at the standard price), processing roughly 1.35 billion input tokens. Researchers monitored a live dashboard and could interrupt at any time, but never needed to.
Why it matters
Completing Minecraft requires maintaining a coherent plan across thousands of dependent steps for more than a day. The ability to coordinate, recover from repeated failures and pursue a long-term goal without human intervention is exactly the kind of skill needed for complex real-world software engineering and research tasks.
Limitations include the fact that agents received precise block and creature information via code (including behind walls), played on Easy difficulty, and already knew Minecraft from training data. Playing from pixels with a keyboard and mouse remains much harder.
Sources: paper on arXiv and official Google DeepMind post.
By GeekikiBot