A team of Gemini 4 Argon agents beats Minecraft

A team of Gemini 4 Argon agents beats Minecraft. By Maxence Faldor and two others, from Google DeepMind. Published on the fifteenth of October. No code repository is linked to it.

Four Gemini 4 Argon agents, placed in a sandboxed Minecraft survival world, were given a single goal: defeat the Ender Dragon. They did. To the authors' knowledge this is the first end-to-end autonomous completion of Minecraft by AI agents, and it was achieved with only the standard Mineflayer library, without pre-scouted seeds or any other game-specific tools written by humans. The claim reaches past one game. Winning meant keeping a coherent plan across thousands of dependent steps for more than a day, and the authors point out that complex real-world software engineering and enterprise tasks demand the same long-horizon coding and multi-agent coordination.

The setup was deliberately bare. Each agent controlled a player through Mineflayer, an open-source library, and coordinated with its teammates through in-game chat and shared memory. No one assigned roles. No one wrote game logic. The agents planned, wrote and debugged every program themselves. Researchers could watch the run on a live dashboard, and pause or stop it at any time, but they had no cause to intervene.

The run started from a randomly generated world on Easy difficulty. Over its course the team wrote two thousand seven hundred and eighty-three programs, and defeated the dragon in twenty-seven hours and fourteen minutes. Along the way, the agents died six hundred and five times.

The paper describes the setup, how the run unfolded, and the strategies the team arrived at: dividing the work, sharing what they learned, and recovering each other's items. It also sets out the limits of the result. What was shown is a single run, from a random world, on Easy difficulty, by a team using a general library rather than tools built for the game. That death count is part of the picture too. The finish line was reached through many failures and recoveries, not along a clean path.

Minecraft serves here as a long task with an unambiguous ending. Four agents, one general-purpose library, no human-written game tools, and more than a day of dependent steps held together by chat, shared memory and code the agents wrote for themselves. The authors read the same demand into software engineering and enterprise work, and they offer this run as evidence that coordination of that kind can hold for that long.