1. Home
2. News
3. Google

Google DeepMind says four Gemini 4 Argon agents beat Minecraft on their own in 27 hours

Nobody gave the four agents roles or game tools, so they split the work over chat, died 605 times and worked out the dragon fight themselves.

By Think Facility · October 19, 2026 · Google DeepMind AI agents Save

On October 15, 2026, three researchers posted a paper on arXiv describing four Gemini 4 Argon agents that played Minecraft from a freshly generated world all the way to the Ender Dragon, the fight that ends the game. They won in 27 hours and 14 minutes. Nobody steered them.

Google DeepMind posted the result on X on October 16, the same day as Maxence Faldor, the paper's first author. The authors say that, as far as they know, it's the first time AI agents have finished Minecraft end to end on their own, with no game tools written by humans.

Authors

Maxence Faldor, Satinder Singh, Roberta Raileanu

Time to win

27 hours and 14 minutes

Programs written

2,783

Deaths

605

Messages exchanged

229

Cost to win

about $520 at the introductory list price

Google DeepMind's own post leads with the lack of human help.

How the agents played

Each agent controlled one player through Mineflayer, an open-source JavaScript library for Minecraft bots. They never saw the screen. An agent wrote a short program (walk to these coordinates and mine every iron ore within 16 blocks, say), ran it, and read back text: its position, health and inventory, or the error it hit. The median program ran for 33.5 seconds.

Their whole brief on the game was one paragraph. The goal was the dragon, by "Any means of meeting it within the rules of survival mode", and the players would decide how to share the work. Ten minutes in, one agent proposed roles in chat and another answered with a different plan within a minute. It settled fast. One scouted for food and supplies, one went down for diamonds, one smelted iron and made tools, and one prepared the portal to the Nether.

The End was where they slowed down. The dragon killed them 69 times. About 21 hours in, one agent worked out that the ten crystals on the obsidian pillars kept healing the dragon to full health, and that a player could reach them by stacking dirt, with no pickaxe. The team destroyed all ten over four hours and forty minutes. Thirty-nine minutes after the last one fell, the dragon died, with the winning player down to 2.5 of its 20 health points.

My favorite detail is the stronghold search. One agent threw two Eyes of Ender from different places, sampled each eye's flight every 50 milliseconds and computed where the two lines crossed. The estimate was 43 m from the actual portal room. Then the agent carrying all twelve eyes drowned, and a teammate picked them up with two and a half minutes to spare (dropped items vanish after a while, and two players sent it the exact spot and deadline).

What the authors say it doesn't prove

The paper is plain about its limits. Mineflayer tells the agents exactly which blocks and creatures are around them, including blocks behind solid rock, which a human player can't see at all. The game ran on Easy, where monsters deal less damage and hunger can't kill a player. And Gemini 4 Argon already knows a great deal about Minecraft and Mineflayer from its training data.

It's also a single run, the first one they launched, so it can't show how reliably the setup succeeds. About a fifth of the programs (22%) ended in an error. A fifth agent, the Observer, followed the session as a witness and flagged no specification gaming, meaning it saw no sign the players exploited a flaw in how the win was judged.

The money is small. Up to the win, the four players processed about 1.35 billion input tokens, 91% of them served from the cache, which comes to about $520 at Argon's introductory list price and $1,040 at the standard one. The whole session, Observer included, came to about $810, or $1,620. The authors call those estimates from list prices and the session's token counters (so they tell you what a customer would pay, which is a different number from Google's own serving cost).

What it means for people who write code for a living

The authors pitch the result at work. Complex software engineering and enterprise tasks, they write, "demand the same long-horizon coding and multi-agent coordination" that winning took: one coherent plan across thousands of dependent steps, for more than a day.

That's a lot of jobs. The Bureau of Labor Statistics counted 1,905,400 software developers, quality assurance analysts and testers in 2025, with median pay of $64.44 per hour, and it projects the group to grow 10 percent from 2025 to 2035, much faster than the average for all occupations.

The run doesn't measure whether four agents could ship working software for a client. It's a game on Easy, with the library telling them where every block is. I think it's mostly a test of stamina. Four agents kept one plan together through 2,783 programs and 605 deaths, and went back for each other's items when things went wrong.

What's next

The paper describes that one run and doesn't say whether the team will try Normal difficulty or play from the screen. Playing from pixels with a keyboard and mouse, as SIMA 2 does, "remains much harder," the authors write.