Researchers from Google DeepMind used four Gemini 4 Argon AI agents to complete the game Minecraft. The agents defeated the final boss, the Ender Dragon, after more than 27 hours of autonomous collaboration in a survival world set to Easy difficulty.
During the process, the agents wrote and executed 2,783 JavaScript programs and exchanged 229 messages. They died 605 times and generated approximately 63,313 lines of code. The system processed 1.35 billion input tokens and 16 million output tokens across 24,000 model calls.
The agents utilized the Mineflayer library to interact with the game. They created 42 reusable helper functions to automate tasks like resource gathering and movement. One agent successfully located a hidden portal room within 43 meters of its actual position.
The cost of the experiment was approximately $520 based on introductory API pricing, with standard pricing estimated at $1,040. Researchers noted that the agents operated without human intervention, predefined strategies, or custom gaming tools.
The study aims to demonstrate how general-purpose AI can manage complex, multi-step tasks. However, researchers noted that the experiment relied on programmatic access to game data, and it remains unclear if these capabilities translate to real-world environments.
