Four Gemini 4 agents beat Minecraft in 27 hours

Researchers at Google DeepMind placed four agents running Gemini 4 Argon in a fresh, randomly generated Minecraft survival world and told them only to defeat the Ender Dragon (arXiv). The agents did not see pixels or press keys. Each controlled a player through Mineflayer, an open-source JavaScript library for Minecraft bots, by writing a short program, running it, and reading back text describing its position, health, inventory, findings or errors. Nobody assigned roles, supplied game strategy, or wrote helper tools; each was the same general-purpose coding agent, with the tools it uses for software work and nothing specific to Minecraft.

From the first program to the dragon's defeat took 27 hours and 14 minutes, 2,783 programs, 229 chat messages and 605 deaths. The agents split up the work within the first quarter hour, posted coordinates and lessons to a shared memory, recovered each other's items after deaths, and turned failures into team rules. One agent located the underground stronghold by throwing two Eyes of Ender, sampling their flight and computing where the two lines crossed. The fight in the End took nearly 14 hours, and the dragon killed the players 69 times. The turn came about 21 hours into the run, when an agent worked out that the crystals healing the dragon could be reached without a pickaxe by stacking dirt beside their pillars; the team destroyed all ten over the next four hours and forty minutes, and the dragon died 39 minutes after the last one. A fifth Observer agent watched for cheating and found none, and the win is recorded in the server log. The authors estimate the four players' model usage up to the win at about $520 at Gemini 4 Argon's introductory list price.

The paper is explicit about what this does not show. Mineflayer hands the agents exact information about surrounding blocks, including ones hidden behind rock, and even the dragon's internal state, and its pathfinding plugin plans routes; playing from pixels with a keyboard and mouse remains much harder. The game was on Easy, the model already knows a lot about Minecraft and Mineflayer from training, 22% of programs ended in errors, and this was a single run, so it says nothing about how reliably the setup succeeds. What distinguishes it from two earlier community dragon kills, according to the authors, is that the agents received no game-specific tools, curriculum or route. Ethan Mollick, sharing the paper, pointed to it as an accessible example of agents coordinating on their own toward a complicated goal.