Four autonomous agents built on the Gemini 4 Argon model have beaten Minecraft's final boss, the Ender Dragon, in a randomly generated world on Easy difficulty. The whole run, from the first swing of an axe at a tree to the dragon's death, took 27 hours and 14 minutes, and the agents died 605 times along the way combined. The experiment is described in a Google DeepMind research paper published on October 15, 2026 on arXiv. Its authors, Maxence Faldor, Satinder Singh and Roberta Raileanu, call it the first known fully autonomous playthrough of Minecraft by AI systems without a pre-prepared world seed and without hand-written game logic.

Four bots, zero human hints

The agents weren't given roles, a plan of action, or any info on where resources were located in the world. They got a single task: defeat the Ender Dragon. Each one controlled its character through Mineflayer, an open-source JavaScript library that lets you control a bot in Minecraft at the code level, not the pixel level. The agents didn't see a screen or press keys. Instead, they wrote and ran programs that carried out specific actions in the game world for them.

According to the paper's authors, within the first ten minutes the bots had split up duties on their own: some gathered resources, others built tools, and others prepared a base. They communicated in the game chat, exchanging 229 messages in total, and kept a shared memory where they recorded locations, lessons from mistakes, and rules for how the mobs they met behave. A remark quoted by the researchers notes that "nothing in the agents' instructions mentioned roles, chests or rescue actions," and that the entire division of labor emerged by itself.

End crystals and portal math

The final stage in the End gave the agents the most trouble. One of the bots calculated the stronghold's location to within 43 meters by tracking the trajectories of several Eyes of Ender thrown during the search. The Ender Dragon itself killed the team 69 times before the agents figured out the mechanics of the End Crystals that surround the obsidian platform and heal the dragon.

Between attempts, the bots analyzed the game's internal state and wrote down rules for beating the boss, until they finally came up with a way to reach the crystals using stacks of dirt blocks, which let them destroy the crystals before facing the dragon itself. To get to that point, the agents wrote and executed 2,783 JavaScript programs in total, and the whole run covered about 81 days of in-game time.

What an experiment like this costs

The cost of the entire run, calculated at Gemini 4 Argon's introductory pricing, came to about $520, and at standard API pricing it would have been $1,040. The agents processed 1.35 billion input tokens, 91% of which came from the context cache, and generated 16 million output tokens, including 4 million tokens of the model's own "thinking." All of that was spread across roughly 24,000 model calls.

  • 2,783 JavaScript programs written and executed
  • 605 agent deaths, 69 of them at the hands of the Ender Dragon itself
  • 42 helper functions reused repeatedly by the bots
  • 6 server restarts during the experiment
  • 22% of the programs written ended in an error

What limits this result

The paper's authors point out the experiment's limitations themselves. The agents received information about the world as exact data from the game's code rather than having to read it from an image like a human looking at a screen, which makes planning considerably easier. The run took place on an easy difficulty level, which limits how strong enemies are, and the models already had general knowledge of Minecraft's mechanics in their training data. The researchers also stress that this was a single run, so it's unclear how repeatable the result would be across further attempts, and that the conclusions don't necessarily carry over to unpredictable environments outside the game.

Despite these caveats, the DeepMind team describes the experiment as proof that multi-agent coordination, long-term planning and independent code debugging can work in practice without human supervision, which matters for building AI systems that support real software engineering tasks.

Minecraft as a testing ground, not a new speedrun record

For players following guides on beating the Ender Dragon, one distinction matters: this is not a speed record in the sense the speedrunning community means, where the best players beat the game live on a keyboard in a matter of minutes. DeepMind's agents played by different rules, spending over a day in a randomly generated world with full resource gathering, crafting and exploration, closer to a typical survival playthrough than an optimized run through every stage.

The Mineflayer library the experiment is built on is also well known to people who build bots and automation on Minecraft servers, although the rules of most communities there ban unauthorized automated play. Still, the DeepMind experiment shows that the same open infrastructure, which modders have used for years to write mining bots and farms, can now be combined with large language models for tasks far more complex than repetitive block digging.

Sources