Four agents powered by Google’s Gemini 4 Argon defeated Minecraft’s Ender Dragon without human intervention, completing a randomly generated game in 27 hours and 14 minutes. The DeepMind team did not assign the agents roles, strategies or game-specific tools.
Instead, the agents divided work on their own, warned one another about threats, recovered items after teammates died and recorded errors in shared memory. During the run, they wrote 2,783 programs, exchanged 229 chat messages and died 605 times. Those numbers show that the successful result depended on extensive trial, software generation and coordination rather than a short sequence of flawless decisions.
The experiment provides a concrete test of agents operating over a long horizon in a dynamic environment. Minecraft requires gathering resources, crafting equipment, navigating and adapting when plans fail, making it harder than a fixed question-and-answer benchmark. The run was also expensive: the reported cost was $1,620 at the standard Gemini 4 Argon price scheduled for 2027. Beating the game demonstrates greater autonomy, but the time, failure count and cost remain important constraints for anyone considering similar multi-agent systems for practical work.