Science

Four Gemini 4 Argon agents beat Minecraft with no human in the loop

DeepMind team needed 27 hours and 605 deaths to kill the Ender Dragon, each conquered benchmark quietly retires so the next one can be sold

Images

The four Gemini 4 Argon agents and their progression through Minecraft. | Image: Faldor et al. The four Gemini 4 Argon agents and their progression through Minecraft. | Image: Faldor et al. Faldor et al.

Four AI agents running Google's Gemini 4 Argon have killed Minecraft's Ender Dragon without a human touching the controls. According to a DeepMind paper, it is the first time the game has been completed with no human in the loop. The run lasted 27 hours and 14 minutes.

The team placed the agents in a randomly generated world with no assigned roles, no strategy and no game-specific tools. The four divided the work themselves, warned each other about threats, recovered items from dead teammates and logged their mistakes in a shared memory. Over the session they wrote 2,783 JavaScript programs, exchanged 229 chat messages and died 605 times.

Minecraft has served as a test of general agency since Microsoft's Project Malmo and OpenAI's MineRL turned it into a benchmark about a decade ago. It belonged to the class of games that demanded purpose-built systems trained from scratch; a general model now clears them out of the box. That is the actual result, and it is quieter than the announcement implies — the run covered exploration, crafting and combat, but the numbers describe persistence more than insight, and 605 deaths is a great deal of retrying. Beating the game with no human input stood as an open challenge for years; the paper that settles it reads like a speedrun log rather than a eureka.

The pattern is older than this run. Every milestone defined as the measure of machine intelligence stops measuring anything on the day it is passed: the lab that set the exam announces the grade, retires the question and enrols the public in the next one. Each capability demo also arrives alongside a compute buildout that needs agent wins to justify its budget, which keeps the announcements coming whether or not anyone outside the field is listening. A goal the field treated as distant for a decade has now been met, and the world outside the labs mostly did not look up.

The price tag carries its own message. The session cost $1,620, calculated at the standard Gemini 4 Argon rate — a rate that does not take effect until 2027.