Gemini 4: Four AI Agents Beat Minecraft in 27 Hours
In a recent experiment, researchers from Google demonstrated that four autonomous AI agents powered by the Gemini 4 Argon model can successfully complete Minecraft. Over the course of 27 hours, the agents navigated a randomly generated Survival world to defeat the Ender Dragon without any human assistance.
Autonomous Coordination and Role Distribution
The experiment relied solely on the open-source Mineflayer library, avoiding the use of pre-known world seeds or human-developed game-specific tools. The agents were given a single high-level goal: defeat the Ender Dragon. To achieve this, they used a shared chat and a common memory to communicate, effectively mimicking the coordination of human players.
As the experiment progressed, the agents autonomously organized themselves into specialized roles. One agent focused on crafting tools, another managed food and supplies, a third handled portal activation, and the fourth constructed defensive structures. A fifth agent, named Observer, acted as an independent monitor to detect any unauthorized assistance and generate live reports.
Computational Scale and Experiment Cost
The journey was far from seamless, resulting in 605 deaths throughout the session. According to Google's researchers, the agents wrote and executed 2,783 programs and exchanged 229 messages. The total computational load involved processing 1.35 billion input tokens.
This scale of operation comes with a significant cost. Based on the researchers' data, Engadget estimated the API costs for the experiment to be between $810 and $1,620, depending on whether the standard or list price for Gemini was applied.
The research team highlights this as the first fully autonomous end-to-end completion of Minecraft by AI agents using only the standard Mineflayer library. The key achievement lies in the agents' ability to follow a cohesive, long-term plan spanning thousands of interdependent steps over more than a day.
Гика
