Four Gemini Agents Beat Minecraft in 27 Hours — After 605 Deaths and 2,783 Self-Written Programs

Four AI agents have done what no general-purpose AI system had done before, according to Google DeepMind: beaten Minecraft start to finish on their own. A paper dated October 15 by Maxence Faldor, Satinder Singh and Roberta Raileanu describes four Gemini 4 Argon agents that defeated the Ender Dragon in 27 hours and 14 minutes, starting from a randomly generated world with no human input during gameplay and no researcher-assigned plan.

The agents did not play by watching a screen. Each controlled one player through Mineflayer, an open-source JavaScript library for Minecraft bots, and wrote its own short programs to walk, mine, craft and fight. Each program returned a text report on position, health, inventory and errors, which the agent used to decide what to code next. The median program ran 33.5 seconds before completing or hitting a self-chosen time limit.

The run's statistics read like a marathon log. The team wrote 2,783 programs totaling roughly 63,313 lines of JavaScript, exchanged 229 chat messages, and died 605 times before winning. Within the first ten minutes they had divided responsibilities — one gathered supplies, another focused on tools and exploration — and they maintained shared notes on locations, strategies and lessons from failures, eventually developing 42 reusable helper functions.

One moment stands out as pure reasoning: an agent estimated the location of a hidden stronghold by tracking the flight paths of two Eyes of Ender and computing where they intersect, landing within 43 meters of the actual portal room. The endgame took persistence — the agents reached the End dimension at 13 hours 33 minutes, but the dragon repeatedly killed them until they identified the ten crystals healing it, bridged up on stacked dirt blocks to destroy them, and finished the fight 40 minutes after the last crystal fell at 26 hours 34 minutes. Six server restarts interrupted the run.

The compute bill was modest by frontier standards: about 1.35 billion input tokens and 16 million output tokens across roughly 24,000 model calls, which the researchers estimate at about $520 at Gemini 4 Argon's introductory API pricing, or $1,040 at standard rates. A fifth agent verified claims against the run record, and the researchers report no safety incidents, while retaining the ability to pause or stop the experiment.

DeepMind calls it the first known autonomous completion of Minecraft using general-purpose agents without custom gaming tools or predetermined world configurations. But the paper is careful about scope: the result demonstrates sustained coordination inside a controlled sandbox, not reliability on software engineering or workplace tasks, where goals are messier and mistakes cost more than a respawn.