DeepMind has four Gemini agents team up to defeat the Ender Dragon in Minecraft

How to chop trees, mine, craft equipment, and defeat the Ender Dragon—all of this became research content. Once the research is done, they publish papers and compete against other AI systems.

Some have a single Claude AI take on the challenge, some study speedrunning with Jev, and recently Google had four Geminis team up and "journey to the west" starting from zero.

From "going it alone" to "the four disciples"

Back then with Claude, developers prepared tools for it to mine, craft equipment, and fight, letting it decide what to do next and then call the tools itself to act. In the end, Claude took less than four hours to defeat the game's big boss, the Ender Dragon.

Claude Opus 5 defeats the Ender Dragon in under four hours | YouTube

A more advanced version. Two models cooperate: GPT6-Astra handles planning, and Jev (good at fast decision-making) chooses actions, clearing the game in eight minutes and forty-three seconds. At the time this was touted as a "Jev success story" and went viral on social media for a while.

The latest development is four Gemini 4 Argon agents teaming up to build Minecraft from scratch by hand, with only one fixed goal in mind: defeat the boss.

The four Geminis enter the game empty-handed, with no human-written programs to rely on; chopping trees, mining, and crafting equipment all have to be learned from scratch. To chop a tree, write a piece of tree-chopping code; to mine, write a piece of mining code. Write it, run it, and if the result is wrong, rewrite it.

A stumbling journey to completion | DeepMind

The "four-person squad" really is like the Journey to the West.

Before the Ender Dragon, they killed other monsters 351 times. The four AIs drowned 67 times, and for that they even wrote a dedicated program on how to keep the game character floating on the water's surface.

Let's unpack one of these "trials" for you: at one point, one AI was carrying key materials needed to fight the dragon, was still eight meters from its teammates, fell into a river, and drowned. The remaining two AIs quickly sent messages, reported the death location, and urged the nearest teammate to go pick up the key materials. Fortunately, that teammate was also a smart one.

The four agents also share experience with each other | DeepMind

The dragon fight took them nearly fourteen hours, and they were killed by the dragon 69 times. The more they got beaten, the more experience they naturally gained. The AIs specifically reminded each other: don't just stand still when you first enter the battlefield—the dragon will charge over and knock you flying.

Writing and revising all the way, the four roamed the game world for more than eighty days. They died a total of 605 times, wrote more than 2,700 programs, and finally conquered the Ender Dragon after 27 hours of real time.

They also wrote 63,300 lines of code | DeepMind

This long expedition condensed into DeepMind's paper—research on how AI can defeat the ultimate boss without humans playing on its behalf.

In 2016, five computer scientists at Microsoft spent their days researching how to get AI to climb a mountain in Minecraft. They later opened up a research platform called Malmo so developers worldwide could use the game to train AI.

In 2019, MineRL (a Minecraft AI research project initiated by Carnegie Mellon University) launched a competition requiring participants, within a limited training time, to teach AI to mine diamonds on its own.

You might think, isn't it just mining a diamond? It wasn't until 2025 that Google DeepMind, using DreamerV3, first achieved this without relying on human demonstrations or pre-arranged learning curricula, letting AI figure out from scratch how to mine a diamond. This result was even published in Nature.

The 2022 MineDojo project collected over 730,000 YouTube gameplay videos, about 7,000 pages of wiki encyclopedia, and 340,000 Reddit posts. The videos alone totaled roughly 300,000 hours, along with corresponding English commentary text. This massive amount of Minecraft material is also key to AI's ability to improve at this game | MineDojo

In 2022, OpenAI collected 70,000 hours of human gameplay videos to teach AI to play Minecraft via mouse and keyboard like a real person. This approach is called VPT (Video PreTraining): the AI doesn't need developers to tell it in advance which key to press at each step; instead it learns operations from player videos. The AI learned to craft a diamond pickaxe, though the success rate was only 2.5%.

In 2022, Minecraft was already being used by OpenAI for pretraining video models | OpenAI

In 2023, Voyager (an AI autonomous exploration project) connected GPT-4 into the game, letting the AI assign tasks to itself. When the AI learned a skill, it recorded the code and could call it directly next time, developing the ability to "fix skills and form memory."

Minecraft is a game with an extremely high degree of freedom. It doesn't dictate how users play; building houses, farming, raising animals, or just wandering around can keep you playing for a long time. Hardcore players build the "Seven Wonders" in it, hand-build the Potala Palace, and some even use the game's circuit components to build a computer that can actually compute.

Whether fixing a door on a pigpen or hand-building an 8-bit computer, it all comes down to the same thing: the player needs to perform single operations in Minecraft—moving, mining blocks, placing blocks, and crafting.

Minecraft Universe has compiled 3,452 basic tasks for Minecraft, and these tasks can be assembled like Lego bricks into a vast game world.

For example, the big task "defeat the Ender Dragon" can be broken down step by step into specific smaller tasks.

AI has to prepare for battle—which means preparing weapons—making weapons requires finding iron—finding iron requires mining—mining requires a pickaxe—making a pickaxe. After mining, you also need to find fuel and smelt with a furnace. Halfway through, if you're tired and hungry, you need to eat and sleep. A small task like "making weapons" spins off an entire production line.

Gemini's own understanding of the task chain | DeepMind

Many tests for AI only look at whether it can give the standard answer. But Minecraft is very well suited to testing AI's "subjective initiative": what to do first, how many materials are enough, which teammate is reliable and suited for what—all of this the "player" has to arrange itself (write programs and test whether the programs run).

Of course, the premise is that the game provides enough openness. The game map can be randomly generated, and experiments can be reproduced using the same map seed. Resources, monsters, day and night, and game difficulty can all be adjusted. Although it's just a bunch of blocks, it extends to and encompasses elements like space, resources, tools, and environmental change.

Creativity, communication, trial and error, summarization, and collaboration—aren't they all trained here | DeepMind

Today, humans "directing" AI has gone from "translate this paper" to "build me a piece of software." Writing a short piece of code is no longer rare, but making them work continuously for over ten hours, calling tools thousands of times in between, having memory, understanding cooperation, and being able to coordinate—that's enough to keep them busy for a while.

So compared with speedrunning the game, the stumbling pioneering gives scholars richer research material.

In 2016, AlphaGo defeated Lee Sedol, and one game of Go introduced "artificial intelligence" to the whole world. In 2026, four agents defeated the Ender Dragon. A different game, and perhaps also a new starting point.

References

This article comes from the WeChat public account "Guokr" (ID: Guokr42), author: Gaoji Dongwu, published by 36Kr with authorization.