Google DeepMind AI Agents Beat Minecraft in 27 Hours With No Human Help (2026)

Four AI agents powered by Google’s new Gemini 4 Argon model have beaten Minecraft from start to finish with no human intervention, taking 27 hours and 14 minutes, writing 2,783 programs and dying 605 times along the way.

Official Minecraft key art — Image: Mojang Studios

In a research paper posted to the arXiv preprint server on October 15, 2026, Google DeepMind researchers Maxence Faldor, Satinder Singh and Roberta Raileanu describe dropping four general-purpose coding agents into a randomly generated Minecraft survival world on Easy difficulty with a single instruction: defeat the Ender Dragon. Google DeepMind announced the result publicly on October 16.

Each agent controlled its own in-game player through Mineflayer, an open-source library that lets code drive a Minecraft character, and the four coordinated through in-game chat and a shared memory store. Nobody assigned roles, nobody wrote game strategy, and no game-specific tools were built by humans. The agents planned, wrote and fixed every program themselves while researchers watched on a live dashboard.

What happened next looked surprisingly like teamwork. Within the first ten minutes, the agents split up the work on their own. They warned each other about threats, shared coordinates and lessons, recovered gear dropped by fallen teammates and debugged each other’s errors. Roles emerged naturally: one agent, nicknamed “janesick,” focused on tools, “gyldenia” gathered food and supplies, “auratian” handled the Nether portal, and “bulicame” built shelters for farming Endermen and landed the final blow on the dragon. A fifth agent, called the Observer, watched the entire run to verify nobody cheated and to log events as they happened.

The run was messy, which is part of the point. The agents died 605 times, mostly to monsters, and 22 percent of their 2,783 programs ended in an error. The endgame was the slowest part: the team built a Nether portal inside the first two hours and reached the End dimension after about 13 and a half hours, but the final dragon fight alone dragged on for nearly 14 more hours. In total the session consumed 1.35 billion input tokens, costing about $810 at Gemini’s list price or $1,620 at standard pricing.

This is a different kind of milestone from DeepMind’s famous gaming victories. AlphaGo and AlphaStar, which conquered Go and StarCraft II, relied on systems custom-trained for a single title. The Minecraft agents were general-purpose coding assistants, the same kind of software engineers use for daily programming work, pointed at a game instead of a codebase.

The researchers say that is exactly why the experiment matters. Beating Minecraft required holding one plan together across thousands of dependent steps for more than a day, recovering from failures and coordinating with teammates throughout. Those are the same skills AI agents need for real software and enterprise work, where progress depends on completing many steps in the right order. The paper frames Minecraft as a public, verifiable test of long-horizon teamwork.

There are real caveats, and the researchers are upfront about them. The game ran on Easy difficulty, so monsters dealt less damage. The agents did not play from pixels like a human would; Mineflayer handed them structured information about the world, including blocks hidden behind solid rock. The model already knew Minecraft’s optimal strategies from its training data. And this was a single successful run, the first attempt the team tried, so nobody knows how reliably the setup would succeed again. The researchers write that to their knowledge this is the first end-to-end autonomous completion of Minecraft by AI agents using only the standard Mineflayer library, with no pre-scouted seeds or human-written game tools.

It is a long way from beating a blocky dragon to doing real work, and the paper’s limits section makes that clear. But for the first time, a team of AI agents sustained a plan, split the labor and finished the job over more than a day with nobody at the keyboard. That is the part researchers across the industry will be studying.

Sources: TechRepublic | Engadget | INVEN Global | The AI Insider

Axee Davies

Axee Davies is the founder of AxeeTech, publishing gaming guides, codes, and Windows how-tos since March 2013. A tech enthusiast for 13+ years — starting with smartphones and Android rooting — he now runs a site that puts out a new guide every day for Roblox players, traders, and Windows users. More about AxeeTech

axeetech.com