# Four Google DeepMind AI agents beat Minecraft in 27 hours > Source: > Published: 2026-10-19 16:16:22+00:00 Four agents built on Google's Gemini 4 Argon model beat Minecraft with no human playing. Twenty-seven hours and 14 minutes after they started, they killed the Ender Dragon, the game's final boss. The researchers at Google DeepMind who set up the run were not asking whether a machine can play a game. They wanted to know whether it can hold one plan together for more than a day, which is what long software and business jobs ask of it, and what models have struggled to do. ## What the four agents actually did The four agents began in a [randomly generated Minecraft survival world](https://arxiv.org/abs/2610.22319v1) with a single instruction: defeat the Ender Dragon. No one assigned roles, no one wrote game strategy for them, and the world was not one a human had scouted in advance. They did not watch the screen or press keys as a person would. Each agent controlled its character through [Mineflayer, an open-source library](https://arxiv.org/html/2610.22319v1) that lets a program drive a Minecraft player, and worked by writing short JavaScript programs: walk here, mine this, craft that. It then read text back about where it was, how much health it had and what it was carrying, and wrote the next program from there. Across the run, the four wrote and ran 2,783 programs, 63,313 lines of code in all. ## Why the game is the easy part A dragon fight is not what makes this hard. Reaching the dragon takes a long chain of ordinary steps, each of which assumes the last one worked: gather wood, stone, iron and food, make tools, find a rare underground portal, fill it with the items that open the way, then survive the fight. Fail anywhere and the plan has to survive the failure. The authors describe the difficulty as one of horizon. Holding a single coherent plan together across [thousands of dependent steps, for more than a day](https://arxiv.org/abs/2610.22319v1), is closer to what complex software and business work demands than a short puzzle is. Their numbers show how little of it went smoothly: the team died 605 times, [69 of those deaths at the hands of the dragon](https://arxiv.org/html/2610.22319v1), and 22% of the programs they wrote ended in an error. ## They organised themselves None of the division of labour was designed by the researchers. The agents proposed jobs to each other in the game's chat: one took iron and smelting, another wood, food and flint, others cave and lava exploration for diamonds and obsidian. They sent [229 messages before the win](https://arxiv.org/html/2610.22319v1), and kept notes in a shared file so a lesson learned by one agent survived the death of the agent who learned it. The clearest example is a rescue. One agent, auratian, drowned while carrying all twelve Eyes of Ender, the items the team needed to open the way to the dragon. It died about eight metres from a teammate. The others posted the exact spot and the deadline before the dropped items would disappear, and bulicame picked them up with roughly two and a half minutes to spare. ## What it cost The report puts the run up to the win at [about $520 at Gemini 4 Argon's introductory list price](https://arxiv.org/html/2610.22319v1), or $1,040 at the standard price. The whole session, which included the monitoring agent and a further day of activity after the win, came to about $810, or $1,620 at standard prices, over roughly 1.35 billion input tokens and 16 million output tokens. For scale, [Google's published price for the model](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon) is $2 per million input tokens and $10 per million output tokens at the introductory rate, rising to $4 and $20. ## How the result was checked A fifth agent, called the Observer, followed the run without playing. It acted as [an independent witness](https://arxiv.org/html/2610.22319v1) and reported no specification gaming, the failure mode where a program meets the letter of its goal in a way nobody intended, and it never joined the game or ran a program in it. ## The limits the authors state The claim to be first carries the authors' own qualifier. In the report, they wrote that this is the [first end-to-end autonomous completion of Minecraft by AI agents "to our knowledge"](https://arxiv.org/abs/2610.22319v1), achieved with only the standard Mineflayer library and without pre-scouted seeds or other game-specific tools written by humans. The agents also played on Easy difficulty, and they acted through code rather than a screen, which is not how a person plays. Roberta Raileanu, one of the three authors, wrote: ["An independent research experiment, not an official Minecraft milestone, and not approved by or associated with Mojang or Microsoft."](https://x.com/robertarail/status/2111027798290600136) Maxence Faldor, the first author, posted: ["Four Gemini 4 Argon agents, one shared Minecraft world, one goal: defeat the Ender Dragon. 27 hours and 2,783 self-written programs later, the dragon fell."](https://x.com/maxencefaldor/status/2111023397106770026) Those caveats narrow the claim without erasing it. Nothing in the run was handed to the agents except the goal, and the plan had to survive 27 hours of their own mistakes in a world no one had prepared. ## Where the same model is already working The clearest sign of what the experiment is for sits outside the game. Google says Gemini 4 Argon agents are already [moving the company's C and C++ code to Rust](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon), including work on the Fuchsia operating system kernel measured in up to 800,000 lines, and that internal agents freed more than 300 tebibytes of data-centre memory, with savings estimated at between 500 tebibytes and a petabyte. Those are Google's figures for its own systems, not independent measurements. The run itself settles something narrow: a general-purpose coding model, given no game-specific help, kept one plan alive for more than a day in a world it had never seen. The jobs the authors point to run for weeks, in environments that were not built to be survived, and Google's own account of Argon agents migrating its codebase is the nearest the evidence comes to that test, measured by the company rather than by anyone outside it.