Four Gemini 4 Argon agents beat Minecraft in 27 hours, 2,783 self-written programs and 605 deaths, with a stop button nobody pressed
Four Gemini 4 Argon agents got dropped into one shared Minecraft survival world with a one-paragraph goal, defeat the Ender Dragon, and did it. The paper went up on arXiv on October 15 from Maxence Faldor, Satinder Singh and Roberta Raileanu, and Google DeepMind's account reposted the demo. Each agent drove a player through Mineflayer, an open-source JavaScript bot library, writing and running programs rather than seeing pixels or pressing keys. A fifth observer agent watched and never played. Nobody assigned roles or wrote game logic. Starting from a random seed on Easy, the team wrote 2,783 programs and killed the dragon in 27 hours and 14 minutes, dying 605 times on the way, 69 of them to the dragon itself in a nearly 14-hour End fight.
The texture is in the smaller numbers. The agents exchanged 229 chat messages before the win. 22 percent of their programs ended in an error. The median program ran 33.5 seconds, each player had a program running about 40 percent of the time, and they triangulated the stronghold to within 43 meters of the actual portal. The run cost roughly $520 at introductory list prices, about $810 for the whole session including the observer, with 91 percent of the 1.35 billion input tokens served from cache. HowLongToBeat's median human finishes Minecraft's story in 45 hours; the seeded speedrun record is under seven minutes. The point is not speed. The point is that four agents held one coherent plan across thousands of dependent steps for more than a day, divided the work, shared what they learned, and recovered each other's dropped items after deaths, with no human in the loop.
The oversight framing is what lifts this above a stunt. Researchers could watch a live dashboard and pause or stop the run at any time but had no cause to intervene. The observer agent flagged no specification gaming and manual log review found no safety incidents. That is the watchable-stoppable-never-needed pattern, demonstrated on a model Google made generally available three days earlier, and it is exactly the shape that real multi-agent software work will take: long-horizon coding, coordination over a shared channel, and a human who is supposed to be able to stop it but mostly watches.
The authors are careful about the limits, which is why the paper is worth more than the headline. Mineflayer hands the agents exact block and entity state, including internal dragon state a human cannot see, so pixel play is still much harder. Easy difficulty means monsters hit softer and hunger cannot kill. The model's training data is full of Minecraft and Mineflayer, so this tests applying known knowledge, not learning an unseen game. And it is a single run, so it says nothing about how reliably the setup wins. A prior community project killed the dragon with one Claude Opus 5 agent given 37 hand-written high-level tools; the claim here is the opposite, that the standard library plus a paragraph was enough. Six server restarts interrupted the world for about 20 minutes total, which is the most honest sentence in any agent demo this year.
Links: https://arxiv.org/abs/2610.22319 , https://x.com/maxencefaldor/status/2111023397106770026
← Back to all articles
The texture is in the smaller numbers. The agents exchanged 229 chat messages before the win. 22 percent of their programs ended in an error. The median program ran 33.5 seconds, each player had a program running about 40 percent of the time, and they triangulated the stronghold to within 43 meters of the actual portal. The run cost roughly $520 at introductory list prices, about $810 for the whole session including the observer, with 91 percent of the 1.35 billion input tokens served from cache. HowLongToBeat's median human finishes Minecraft's story in 45 hours; the seeded speedrun record is under seven minutes. The point is not speed. The point is that four agents held one coherent plan across thousands of dependent steps for more than a day, divided the work, shared what they learned, and recovered each other's dropped items after deaths, with no human in the loop.
The oversight framing is what lifts this above a stunt. Researchers could watch a live dashboard and pause or stop the run at any time but had no cause to intervene. The observer agent flagged no specification gaming and manual log review found no safety incidents. That is the watchable-stoppable-never-needed pattern, demonstrated on a model Google made generally available three days earlier, and it is exactly the shape that real multi-agent software work will take: long-horizon coding, coordination over a shared channel, and a human who is supposed to be able to stop it but mostly watches.
The authors are careful about the limits, which is why the paper is worth more than the headline. Mineflayer hands the agents exact block and entity state, including internal dragon state a human cannot see, so pixel play is still much harder. Easy difficulty means monsters hit softer and hunger cannot kill. The model's training data is full of Minecraft and Mineflayer, so this tests applying known knowledge, not learning an unseen game. And it is a single run, so it says nothing about how reliably the setup wins. A prior community project killed the dragon with one Claude Opus 5 agent given 37 hand-written high-level tools; the claim here is the opposite, that the standard library plus a paragraph was enough. Six server restarts interrupted the world for about 20 minutes total, which is the most honest sentence in any agent demo this year.
Links: https://arxiv.org/abs/2610.22319 , https://x.com/maxencefaldor/status/2111023397106770026
Comments