←News detail

The PaperUpdated Oct 24 12:41

Four Gemini agents autonomously complete Minecraft, defeating the Ender Dragon in 27 hours

AI summary

Four autonomous agents running Gemini 4 Argon took 27 hours and 14 minutes to defeat the Ender Dragon in Minecraft survival mode, with no human intervention throughout and without receiving game-specific training; the DeepMind research team disclosed this experiment in an arXiv paper. During the process, the agents wrote and ran 2,783 programs, exchanged 229 chat messages, died 605 times, processed about 1.35 billion input tokens, and cost about $520 at discounted prices. The paper states this is the first end-to-end autonomous completion using only the standard Mineflayer library, without pre-exploring seeds or human-written game-specific tools.

Why read this

SelectedAI score 80

The paper discloses that four general-purpose agents autonomously completed the game without game-specific training, comparable to the specialized training approaches of AlphaGo and AlphaStar.

Topics

Tags

Body · AI translation

Download the app

+1 Four Gemini agents autonomously complete Minecraft, defeating the Ender Dragon in 27 hours The Paper reporter Ji Jingjie 2026-10-22 11:30 Source: The Paper ∙ Future 2% >Listen to full textFont size

Four autonomous agents running Gemini 4 Argon defeated the final boss, the Ender Dragon, in the survival mode of the game Minecraft in 27 hours and 14 minutes, without any human intervention and without receiving any game-specific training. The DeepMind research team under Google recently disclosed this experiment in a paper published on arXiv.

In the paper, the researchers describe this achievement as "to our knowledge, the first end-to-end autonomous completion by an AI agent using only the standard Mineflayer library, without pre-explored seeds or human-written game-specific tools."

This constitutes a methodological departure from DeepMind's previous game AI achievements. When AlphaGo defeated Lee Sedol in 2016, its core was the combination of deep reinforcement learning and tree search, and the system was specifically trained for Go using human game records and self-play. In 2019, AlphaStar achieved a 10:1 win rate against top human players in StarCraft II, also relying on reinforcement learning training specific to that game.

Gemini 4 Argon, by contrast, is a general-purpose model; in the experiment, apart from the general-purpose robotics library Mineflayer, there was no logic tailored for Minecraft.

Maxence Faldor, one of the paper's authors, said on social media that DeepMind plans to apply the same setup to programming, science, and robotics tasks.

Throughout the process, the four agents wrote and ran 2,783 programs, exchanged 229 chat messages, and died 605 times, 69 of which were at the hands of the Ender Dragon.

The experiment prohibited the assistance commonly seen in previous game AI achievements. The four agents were placed into a randomly generated survival-mode world with the difficulty set to "Easy" and cheat commands disabled. No one assigned them roles, formulated strategies, or provided tools specific to Minecraft. Each agent controlled a game character through the open-source Mineflayer library, acting by writing and running short JavaScript programs—for example, commanding the character to walk somewhere and mine nearby iron ore—and then reading back the returned position, health, and inventory information, or reading program errors.

At about 21 hours into the game, one of the agents discovered that the ten crystals atop the pillars were continuously healing the Ender Dragon, and found a way to reach the crystals by stacking dirt blocks. Over the next 4 hours and 40 minutes, the four agents destroyed all ten crystals, and 39 minutes later, one agent team member killed the Ender Dragon with only 2.5 health points remaining (out of a maximum of 20).

In terms of cost, by the time of completion, the agents had processed about 1.35 billion input tokens, which at Gemini 4 Argon's discounted rate comes to about $520, or about $1,040 at the standard rate. The full session, including a fifth "observer" agent responsible for monitoring cheating behavior, cost about $810.

There are still essential differences between agents playing games and humans. The paper notes that Mineflayer provided the agents with precise information about surrounding blocks and monsters, including what was behind solid rock, and that a separate plugin handled path planning. In addition, about 22% of the programs ended in errors.

The paper also notes that two community projects had previously reported earlier Ender Dragon kills, but they used game-specific tools or were completed under fixed, simpler settings.

References:

Source

The Paper