Four Google AI agents beat Minecraft on their own, dying 605 times along the way

Four Google DeepMind Gemini 4 Argon agents autonomously completed Minecraft, killing the Ender Dragon 27 hours and 14 minutes after starting bare-handed in a randomly generated world, according to a paper by researchers Maxence Faldor, Satinder Singh and Roberta Raileanu posted to arXiv on October 15. The agents, which controlled players through the open-source Mineflayer JavaScript library by writing and running short programs, wrote 2,783 programs, exchanged 229 chat messages and died 605 times, with the End fight alone taking nearly 14 hours. The run consumed about 1.35 billion input tokens (91% cached) and 16 million output tokens across roughly 24,000 model calls, costing an estimated $520 at Gemini 4 Argon's introductory list price of $2 per million input tokens and $10 per million output tokens, or about $1,620 for the whole session at standard pricing.

A fan dressed as Minecraft's Steve at Comiket in Tokyo, August 2013 illustrative . Image: Guilhem Vellut <https://commons.wikimedia.org/wiki/File:Minecraft Cosplay @ Comiket 84 - Summer 2013 @ Tokyo Big Sight 9490347671 .jpg / Wikimedia Commons, CC BY 2.0 https://creativecommons.org/licenses/by/2.0/ , cropped Four AI agents from Google DeepMind have beaten Minecraft with nobody telling them how. In a paper posted to arXiv https://arxiv.org/abs/2610.22319 on October 15, researchers Maxence Faldor, Satinder Singh and Roberta Raileanu say a team of Gemini 4 Argon agents started in a randomly generated world with bare hands and killed the Ender Dragon, the game's final boss, 27 hours and 14 minutes later. Along the way the four players wrote 2,783 programs, swapped 229 chat messages and died 605 times. The researchers call it, to their knowledge, the first end-to-end autonomous completion of Minecraft by AI agents using only a standard, open-source bot library, with no pre-scouted world, no human-written game tools and no strategy handed to them. One paragraph of instructions Each agent controlled a player through Mineflayer, an open-source JavaScript library for Minecraft bots. Instead of watching a screen, an agent wrote a short program walk here, mine every iron ore nearby , ran it, read back what happened, then wrote the next one. The only game instruction was a single paragraph: defeat the Ender Dragon, by any means within the rules of survival mode, with the players deciding how to share the work. Cheat commands were off, the game ran on Easy difficulty and no human touched it during the run. A fifth Gemini agent watched for cheating and found none. Ten minutes in, one agent proposed a division of labour in the chat. The team settled into roles on its own: one scouted for food and beds, one dug for diamonds, one built tools and tracked down a Nether fortress and the stronghold. They had iron tools, diamonds and a Nether portal within two hours, and reached the End, the dragon's realm, 13 hours and 33 minutes in. Fourteen hours to kill the dragon The last stretch was the hardest. The fight in the End took nearly 14 hours, because the ten crystals on top of tall obsidian pillars kept healing the dragon back to full health. About 21 hours in, one agent worked out that no pickaxe was needed to reach them: stack dirt beside each pillar, punch the crystal, dig back down. Its message to the team read, in capitals, "NO PICKAXE OR WOOD NEEDED FOR CRYSTALS " The team destroyed all ten over the next four hours and forty minutes, and the dragon fell soon after. The researchers say the team arrived at habits you would expect from a good human team: Nothing in the agents' instructions mentioned roles, chests or rescues, beyond the line that the players decide how to share the work. Maxence Faldor, Satinder Singh and Roberta Raileanu, Google DeepMind About $1,620 for the whole session Up to the win, the four players used about 1.35 billion input tokens 91% served from cache and 16 million output tokens over roughly 24,000 model calls. At Gemini 4 Argon's introductory list price that comes to about $520, or $1,040 at the standard price. The whole session, including the Observer and a day of play after the win, came to about $810, or $1,620 at the standard price. These are estimates from list prices, not Google's own costs. Google's introductory Argon pricing of $2 per million input tokens and $10 per million output tokens runs until November 30 https://madrobot.blog/2026/10/13/google-gemini-4-argon-available-ai-pro-ultra-api-robotics/ , after which it doubles to $4 and $20. Not the first dragon kill, and not from pixels Earlier Minecraft AIs such as OpenAI's Video PreTraining and DeepMind's DreamerV3 were trained on huge amounts of gameplay just to reach diamonds, and agents such as Voyager leaned on Minecraft-specific skill libraries. The paper notes two community projects have recently reported dragon kills too: a single Claude Opus 5 agent using 37 high-level Minecraft tools on Normal difficulty, and a planner on a fixed speedrun world in Peaceful mode. DeepMind's claim rests on what its agents were not given. The authors are open about the limits. Mineflayer tells the agents exactly which blocks and creatures are around them, even behind solid rock, and plans routes for them: far easier than playing from pixels. Easy mode means weaker monsters. Gemini already knows a lot about Minecraft from its training data, 22% of programs ended in an error, and this was a single run, the first one they tried. For scale, the median human takes about 45 hours to finish the main game, according to HowLongToBeat. Why it matters Finishing Minecraft from scratch once needed purpose-built, heavily trained systems; here general coding agents did it by writing their own code and organising themselves over more than a day. That kind of long, self-correcting teamwork is exactly what companies want from AI agents in real software and office work, though one run on Easy mode is a demonstration, not proof that it is reliable. Sources: Faldor, Singh and Raileanu, "A team of Gemini 4 Argon agents beats Minecraft", arXiv https://arxiv.org/abs/2610.22319 ; full paper PDF https://arxiv.org/pdf/2610.22319v1 .