Google DeepMind researchers deployed four AI agents powered by Gemini 4 Argon to complete the game Minecraft. The agents operated autonomously for 27 hours and 14 minutes, successfully defeating the Ender Dragon without human intervention.
The agents communicated via chat to distribute tasks, such as gathering resources, crafting tools, and locating the portal. They utilized a shared memory space to track coordinates, share warnings, and manage custom functions, writing 2,783 programs in JavaScript to execute actions.
The experiment recorded 605 deaths, with 351 caused by monsters, 67 by drowning, and 57 by falling. The agents demonstrated problem-solving skills, such as calculating the location of a stronghold using triangulation and developing strategies to dismantle healing crystals.
A fifth agent, the Observer, monitored the logs to ensure no cheating occurred. The total cost of the experiment, based on API pricing, ranged between $810 and $1,620, depending on the billing tier applied to the session.
DeepMind views this project as a simulation for complex engineering and corporate tasks. The researchers identified three key skills demonstrated by the agents: team coordination, recovery from failure, and long-term goal maintenance.
