In a randomly generated digital wilderness, four artificial minds spent 27 hours learning, failing, and ultimately prevailing — defeating Minecraft's final boss without a single human instruction. Google DeepMind's Gemini 4 Argon agents divided labor spontaneously, recovered from hundreds of deaths, and wrote thousands of small programs to navigate a world they had never seen. The achievement is less about a video game and more about a threshold: machines reasoning through open-ended complexity, adapting to failure, and arriving somewhere no one told them to go.
Google's AI agents autonomously beat Minecraft's final boss in 27 hours
Four agents wrote 2,783 programs and died 605 times to win.
Mark
So these agents actually wrote code to play the game? They didn't just press buttons?
Mimi
Exactly. Each agent generated JavaScript programs—instructions like "walk to these coordinates" or "mine the iron nearby." Then it read back what happened: its new position, health, inventory. If something failed, it got an error message and had to figure out what went wrong.
Luke
But they had perfect vision through walls, right? And a pathfinding plugin doing the navigation work for them?
Mimi
Yes. They could see every block and creature around them, even behind solid rock. That's a huge advantage over a human player. And the pathfinding plugin meant they didn't have to solve navigation from scratch.
Mark
How did they decide who did what? Did someone assign roles?
Mimi
No assignment at all. They just divided the work themselves. One agent became the explorer, another the supply gatherer, one activated the portal, one fought the dragon. They communicated through chat messages.
Luke
How many times did they actually try this before succeeding?
Mimi
The paper reports one successful run. It doesn't say how many failures came before it.
Mark
And the cost—was it expensive?
Mimi
Between $810 and $1,620 for the whole session, depending on which API pricing tier you use. That's processing 1.35 billion tokens.
Luke
They died 605 times. That seems like a lot of failure built into a "success." How much of this was trial and error versus actual planning?
Mimi
The paper doesn't break that down. We know they recovered from deaths, shared notes, and learned—like discovering the healing crystals. But you're right that we don't have a clear picture of how much was systematic problem-solving versus just trying things until something worked.
Mark
What's the real significance here? Is this a big step forward?
Mimi
It's the first time agents have autonomously completed the full game using only standard tools, no custom code. That's notable. But the limitations matter: perfect information, built-in pathfinding, Easy difficulty, a 22% error rate.
Luke
And we're looking at a single run. We don't know if this is repeatable or if they got lucky.
Le Pouls
- Four AI agents were dropped into an unpredictable survival world with one instruction — kill the dragon — and no guidance on how to begin.
- Over 27 hours, the agents died 605 times, generated 2,783 programs, and coordinated through in-game chat to keep their collective effort alive after each setback.
- A critical turning point came 21 hours in, when one agent independently discovered that ten healing crystals were sustaining the dragon — and that stacking dirt blocks could reach them.
- The final blow landed with the last surviving agent at 2.5 health points, marking the first autonomous end-to-end Minecraft completion using only standard, publicly available tools.
- Researchers acknowledged real constraints: perfect environmental visibility, easy difficulty, a 22% program error rate, and a cost of up to $1,620 in API fees for a single successful run.
In a randomly generated digital wilderness, four artificial minds spent 27 hours learning, failing, and ultimately prevailing — defeating Minecraft's final boss without a single human instruction. Google DeepMind's Gemini 4 Argon agents divided labor spontaneously, recovered from hundreds of deaths, and wrote thousands of small programs to navigate a world they had never seen. The achievement is less about a video game and more about a threshold: machines reasoning through open-ended complexity, adapting to failure, and arriving somewhere no one told them to go.
Four AI agents running on Google's Gemini 4 Argon model spent 27 hours and 14 minutes inside a randomly generated Minecraft world and emerged having defeated the Ender Dragon — the game's final boss — without a single human instruction, pre-written strategy, or game-specific tool. Google DeepMind published the findings on arXiv, describing it as the first autonomous, end-to-end Minecraft completion using only standard, publicly available libraries.
The agents began with nothing but an objective. Each controlled a Minecraft character through Mineflayer, an open-source bot library, writing short JavaScript programs to move, mine, and check their surroundings — then reading back results or error messages. Over the course of the run, they generated 2,783 such programs and exchanged 229 messages through the game's chat system, sharing coordinates and hard-won discoveries.
Without being instructed to, the agents divided labor. One explored and engineered, locating the Nether fortress and calculating the stronghold's position by tracking the flight paths of thrown items in code. Another gathered food and supplies. A third activated the Nether portal. The fourth built shelters and, in the end, delivered the killing blow — with just 2.5 health points remaining.
The dragon fight alone consumed nearly 14 hours. The turning point came around hour 21, when one agent discovered that ten crystals atop tall pillars were continuously healing the dragon, and that stacking dirt blocks could reach them. Over the next four hours and forty minutes, all four agents worked together to destroy each crystal. The dragon died 39 minutes after the last one fell.
Researchers were candid about the limitations: the agents had perfect visibility through solid terrain, played on the easiest difficulty setting, and saw 22% of their programs end in errors. The session cost between $810 and $1,620 in API fees. Still, the achievement stands apart from prior community efforts for its use of only standard tools and its navigation of a genuinely unpredictable world.
Four artificial intelligence agents, each running on Google's Gemini 4 Argon model, spent 27 hours and 14 minutes inside a Minecraft world and walked away having defeated the Ender Dragon—the game's final boss—without a single human instruction, strategy assignment, or specialized tool written just for them. Researchers at Google DeepMind published their findings on arXiv, calling it the first time AI agents have autonomously completed Minecraft from start to finish using only standard, publicly available libraries.
The agents began in a randomly generated survival world set to Easy difficulty, with cheat commands disabled. They had one objective and nothing else: kill the dragon. No one told them how to divide labor, what to prioritize, or which direction to walk. Instead, each agent controlled a Minecraft player through Mineflayer, an open-source library designed to let bots interact with the game. When an agent needed to act, it wrote a short JavaScript program—an instruction to walk to a location, mine iron ore, or check its inventory—then read back the result: its position, health status, what it carried, or an error message if something went wrong.
Over the course of the run, the four agents generated 2,783 programs in total. They exchanged 229 messages with each other through the game's chat system, sharing coordinates and lessons learned. They died 605 times. The dragon killed 69 of them.
The agents divided the work without being told to. One agent, named janesick, took on the role of explorer and engineer. It crafted tools, located the Nether fortress, and found the stronghold containing the portal to the End dimension. To pinpoint the stronghold's location, janesick threw two Eyes of Ender from different positions, tracked their flight paths in code, and calculated where those paths would intersect. The calculation placed the portal room roughly 43 meters from the predicted crossing point—a miss of 1.3 kilometers. Another agent, gyldenia, focused on gathering food and supplies to keep the team alive. auratian activated the Nether portal. bulicame built shelters to protect the team from Endermen and, in the final moments, delivered the killing blow.
The agents maintained shared notes, recording important coordinates and discoveries so that when one died, the others could retrieve its items and continue. The actual fight against the dragon consumed nearly 14 hours of the total 27. About 21 hours into the session, auratian made a critical discovery: ten crystals perched atop tall pillars were continuously healing the dragon. The team could reach those crystals by stacking dirt blocks to climb higher. Over the next four hours and 40 minutes, all four agents worked together to destroy each crystal. Thirty-nine minutes after the last crystal fell, the dragon died. bulicame was down to 2.5 health points out of a maximum of 20.
The entire operation processed approximately 1.35 billion input tokens through Gemini 4 Argon's API. At the introductory pricing rate, the researchers estimate the cost at roughly $520; at standard pricing, around $1,040. The full session, including a fifth agent that monitored for cheating and found none, ran between $810 and $1,620 depending on which pricing tier was applied.
The researchers were transparent about the constraints. Mineflayer provides the agents with perfect information about their surroundings—they can see exactly which blocks and creatures exist around them, even through solid rock. A pathfinding plugin handles route planning for them. The agents played on Easy difficulty, the lowest setting. Twenty-two percent of their programs ended in errors. And this result comes from a single successful run; the researchers did not report how many attempts preceded it. Other community projects have reported AI agents defeating the dragon before, but those efforts relied on game-specific tools or played on fixed, pre-designed maps rather than randomly generated worlds. This achievement stands apart for its use of only standard libraries and its navigation of an unpredictable environment.
Citations marquantes
To our knowledge, this is the first end-to-end autonomous completion of Minecraft by AI agents, achieved using only the standard Mineflayer library without pre-scouted seeds or other game-specific tools written by humans.— Maxence Faldor, Satinder Singh, and Roberta Raileanu, Google DeepMind
Envie de l'histoire complète ?
Lire l'original sur TNW ↗
Technology