Gemini Minecraft agents: 4 AI agents beat the Ender Dragon

Maxence Faldor, Satinder Singh and Roberta Raileanu of Google DeepMind describe in a paper on arXiv how four Gemini Minecraft agents built on Gemini 4 Argon beat the game: from a random survival world all the way to killing the Ender Dragon. It took them 27 hours and 14 minutes, with no human help at all. The paper went up on arXiv on October 15, and since October 19 it has been covered by The Next Web, among others, and all over AI X.

What exactly did the Gemini agents do in Minecraft?

The authors write that, to their knowledge, this is “the first end-to-end autonomous completion of Minecraft by AI agents”. The starting conditions were fair: Minecraft Java Edition 1.20.4, a randomly generated world, Easy difficulty, commands disabled, and a single goal: kill the dragon. Nobody assigned roles or wrote game-specific tools for the agents.

Interestingly, the agents never saw the screen. Each one controlled its character through Mineflayer, an open-source library for Minecraft bots. To do anything, an agent wrote a short JavaScript program and then read the result: position, health, inventory or an error. So these are really coding agents that play the game through code.

A few numbers from the paper:

1. 2,783 programs written before the win, 22% of which ended in an error.
2. 229 messages in the in-game chat the agents used to coordinate (plus a shared memory).
3. 605 deaths, 69 of them in the End.
4. About 1.35 billion input tokens (91% from cache) and about 16 million output tokens.

How did the playthrough go?

The agents split the roles among themselves after only about 10 minutes. One (janesick) took on smelting and tools, found a Nether fortress and located the stronghold. The second (gyldenia) gathered food and supplies, the third (auratian) built the Nether portal, and the fourth (bulicame) hunted for diamonds and landed the final blow on the dragon.

The timeline, according to the authors:

1. 1 h 42 min: Nether portal lit, and within 2 hours iron tools and diamonds as well.
2. about 10 h 45 min: stronghold located from two Eye of Ender throws. The estimate was only 43 meters from the portal room.
3. 13 h 17 min: End portal activated.

And then the real ordeal began. The fight in the End lasted almost 14 hours. Only after about 21 hours of play did the agents figure out that the crystals on the pillars were healing the dragon. Destroying all ten took them 4 hours and 40 minutes, and the dragon fell 39 minutes after the last crystal (Minecraft isn’t that easy after all ;)).

How much did it cost?

The authors estimate the cost of the four players at about $520 at Gemini 4 Argon’s promotional pricing, or $1,040 at standard pricing. The whole session, including a fifth agent acting as an observer, came to about $810 ($1,620 at standard pricing). These are estimates based on the price list, not Google’s actual server costs.

That observer didn’t play at all. It watched for the players cheating or “hacking” the task (so-called “specification gaming”) and flagged nothing of the kind. The researchers also had a live view and could stop the run.

Is there a catch?

There are a few, and the authors list them themselves:

1. Mineflayer gives the agents knowledge a human player doesn’t have, e.g. about blocks and mobs behind a wall. On top of that, a pathfinding plugin plans routes for them.
2. Easy difficulty, not Normal or Hard.
3. A single run, so we don’t know how repeatable the result is.
4. The model knows Minecraft from its training data, so its knowledge of the game didn’t come from nowhere.

According to The Next Web, there were already two community projects in which AI killed the dragon, but they relied on tools written for this game or on a fixed, easier world. The paper mentions AI Bengineering (one agent with 37 ready-made tools) and Malde (a planner on a fixed speedrun seed in Peaceful mode).

For me, the most interesting part of this result isn’t Minecraft itself, but the fact that the agents kept a coherent plan across thousands of interdependent steps and more than a day of work. According to The Next Web, Faldor wrote on X that DeepMind wants to use the same setup for coding, science and robotics.

Can we reproduce it?

Not entirely. The paper says nothing about releasing the code or the agents’ logs. The building blocks are available, though: Mineflayer is open source, and Gemini 4 Argon has been available since October 13 to paying Gemini app users and through the API ($2 per million input tokens and $10 per million output tokens during the launch promotion, then $4 and $20).

If we want a feel for how agents that write code in a loop work, it’s easier to start with Claude Code or with DeepSeek for free in OpenCode.

FAQ

Did AI really beat Minecraft on its own?

Yes, in a single run: from a random world to killing the Ender Dragon, with no hints from humans. But on Easy difficulty and with the Mineflayer library, which gives the agents more information than a player sees.

How long did it take the Gemini agents to beat Minecraft?

27 hours and 14 minutes from the first program to victory, almost 14 hours of which went into the fight in the End.

Did the Gemini agents see the game screen?

No. They controlled their characters by writing JavaScript programs and read the results as text.