Four AI agents beat Minecraft's Ender Dragon in a single run

Every Minecraft player knows the path to the End. It runs through a long chain of tasks: gathering raw resources, building tools, finding a hidden underground stronghold, and only then facing the Ender Dragon, the game's final test. Four AI agents have now walked that whole path on their own. In 27 hours and 14 minutes of continuous play, the Gemini 4 Argon agents defeated the dragon, something the researchers say no AI team had done before.

The setup was deliberately spare. The agents received one game-specific instruction, 'defeat the Ender Dragon', and no Minecraft strategy. Unlike many AI experiments, they were not trained on vast quantities of gameplay data; they relied on their existing knowledge and general-purpose coding abilities. They were placed in a shared, randomly generated survival world on Easy difficulty, with cheat commands turned off and a fifth agent acting as observer to check progress and technical errors. Where a human uses a mouse and keyboard, these agents moved their characters by writing JavaScript through Mineflayer, a standard bot control tool, and read text updates about their health, inventory and location.

The road was not smooth. The agents wrote 2,783 custom programs and died 605 times. Nobody assigned them roles: they split mining, crafting and fighting by talking through in-game chat. When an agent died, it returned to the game, learned from the mistake and adjusted both its strategy and its code. 'No one assigned roles or wrote game logic,' the researchers write.

The caveats matter as much as the result. This was a single run, the first launch the team tried, and the researchers acknowledge it cannot show how reliably the setup succeeds. The experiment stays inside one video game with one clear goal.

What changes, and for whom? For now, the gain is for AI research. The result shows that a general-purpose system, given a goal and no playbook, can plan a chain of thirteen steps and correct itself after failures. It does not tell us when such agents will work outside a game, and the researchers offer no timetable for that.