Four Gemini agents beat Minecraft on their own, and died 605 times doing it

Four AI agents just beat Minecraft with nobody steering. Google researchers put four Gemini 4 Argon agents in a randomly generated survival world with one goal, kill the Ender Dragon, and their paper on arXiv says they got it done in 27 hours and 14 minutes. Nobody assigned them roles or wrote any game logic for them.

They didn’t play it the way you or I would. Each agent acted by writing and running a short JavaScript program instead of looking at pixels and pressing keys, and the team wrote 2,783 of those programs along the way. The Decoder reported that they split up the jobs on their own, warned each other about threats and recovered items from dead teammates. They even fact-checked each other. When one agent, janesick, said it was coming back with twelve blaze powder one second after it burned to death, a teammate flagged the contradiction within three minutes. (I’ve been in that group chat.)

The paper is honest about the asterisks. They played on Easy, where hunger can’t kill you, 22% of their programs ended in an error, and the fight in the End took nearly 14 hours. Engadget counted 605 deaths. HowLongToBeat’s median for a person finishing the main story is 45 hours, so that’s quick, but IGN pointed out the speedrun record is six minutes and 36 seconds. The bill came to $810 at the introductory API price, or $1,620 at the standard one.

The part I’d pay attention to has nothing to do with blocks. The researchers say real software and enterprise work needs the same long-horizon coding and teamwork, and here four agents held one plan together for more than a day through hundreds of failures. If you’re handing an agent a job that runs longer than a coffee break, I’d ask how it checks the other agents’ work, because these four did.

The dragon had no idea what was coming. Eventually.