Google DeepMind researchers say a team of four AI agents has beaten the block-building game Minecraft with no human help during the run. AI agents are AI systems that can carry out tasks on their own, such as using apps and websites, instead of only answering questions. Each of the four was a copy of Google's Gemini 4 Argon model controlling one player. Starting in a newly made world on the game's Easy setting, they defeated the Ender Dragon, the win that counts as finishing the game, after 27 hours and 14 minutes. Maxence Faldor, Satinder Singh and Roberta Raileanu describe the run in a paper posted on Oct. 15. The researchers say long jobs at work, like big software projects, need the same skills: sticking to one plan for more than a day through thousands of steps, and working as a team.
The agents never saw the screen or pressed keys. Each one played by writing a short program, such as an order to walk somewhere and dig up iron, running it through Mineflayer, free, open-source software that lets a program control a Minecraft player, and reading the text that came back. Mineflayer can move a player, dig, build and craft, but does not know how to play. Apart from a note that they were a team of four, the agents' only instructions about the game were one paragraph naming the goal. They also had the general instructions they follow on any job: check each step, build reusable tools, write lessons in notes the team shares, and report honestly. Each agent also had the tools it uses for software work, web search included, though the researchers say none of the programs the players ran went online. Researchers could watch on a live dashboard and pause or stop the run, but say they had no reason to step in.
The team organized itself. Ten minutes in, one agent proposed a division of labor in the game's chat, and within two hours the four had iron tools, diamonds and a portal to the Nether, a second world inside the game. There they gathered what they needed to make Eyes of Ender, items that fly toward the nearest stronghold when thrown. One agent found the underground stronghold that holds the portal to the dragon's island by throwing two Eyes of Ender from different spots and working out where their paths crossed; its estimate landed 43 meters from the room that holds the portal. On the dragon's island, the dragon killed the agents' players 69 times. About 21 hours in, one agent noticed that ten crystals on tall obsidian pillars kept healing it, and worked out how to reach them without a pickaxe: stack dirt beside each pillar, climb up and punch the crystal. The team destroyed all ten over the next four hours and forty minutes, and the dragon died 39 minutes after the last one fell. Along the way the agents posted lessons for each other and caught false reports of success, once retracting their own and once flagging a teammate's.
In all, the four wrote 2,783 programs and died 605 times. A fifth copy of the same AI watched as a witness and never played; neither it nor the researchers, who read a sample of the logs, found any cheating. The researchers estimate the four players' work up to the win would cost about $520 at Gemini 4 Argon's introductory public price, or $1,040 at its standard price; those are estimates from list prices, not what it costs Google to run the model. Human players take a median of about 45 hours to finish the game, by one tracking site's count, and the fastest anyone has finished on a random world is under seven minutes. The agents, though, were four players working at once, on Easy, through programs rather than a screen and keyboard.
The result comes with limits the researchers spell out. Mineflayer told the agents exactly which blocks and creatures were around them, even behind solid rock, and even the dragon's internal state, things a human player has to spot by eye or cannot see at all. An add-on to it also planned routes around obstacles. The researchers say an AI that plays from the screen with a keyboard and mouse, as people do, faces a much harder task. On Easy, monsters do less damage and players cannot starve. Gemini 4 Argon already knows a great deal about Minecraft, and about Mineflayer, from the material it learned from, so the run tests whether it can turn that knowledge into working programs and teamwork, not whether it can learn a game it has never seen. About one in five of their programs ended in an error, and the game's server restarted six times. And it was a single run, the first one they tried, so it does not show how often the setup would succeed.
The researchers call it, to their knowledge, the first time AI agents have finished Minecraft on their own with only a standard, published Minecraft bot library, no other game tools written by people and no world picked in advance. They note that two projects by outside developers have reported beating the dragon this year: one with a single agent running Anthropic's Claude Opus 5, which used 37 ready-made Minecraft tools on Normal, a harder setting than this run's Easy, and one on a fixed world chosen in advance, on the game's Peaceful setting. The researchers say every program, chat message and game event was recorded, but the paper does not say whether those records will be released for others to check.