Google's team has published the results of an experiment on the arXiv preprint server in which four autonomous AI agents powered by the Gemini 4 Argon model completely beat Minecraft in survival mode and defeated the final boss, the Ender Dragon, without any human intervention. This is reported by RBC-Ukraine with reference to Google's publication. According to the researchers, this is the world's first completely autonomous Minecraft playthrough by general AI models without a pre-explored world and without human-made algorithms tailored for specific game tasks.
How the Experiment Was Organized
The researchers placed four AI agents into a randomly generated survival world. The models were not assigned specific roles, were not provided with ready-made strategies, and were not equipped with special gaming tools. Each agent controlled a character using the open Mineflayer library and coordinated actions with other participants through the in-game chat and shared memory. The experiment was monitored by a fifth, independent observer agent, who recorded the absence of protocol violations and generated final reports.
Self-Organization Without External Control
A key observation by the researchers was the algorithms' ability to independently distribute responsibilities during the game. Agents warned each other about dangers, shared coordinates of resources and enemies, suggested solutions in difficult situations, and even looted items from the bodies of fallen virtual partners. None of these functions were pre-programmed: the model generated the interaction protocol itself, relying exclusively on the context of the in-game chat and shared memory.
27 Hours of Continuous Work and the Battle with the Final Boss
The total autonomous playthrough process took 27 hours and 14 minutes. A significant portion of this time was spent by the algorithms battling the Ender Dragon, the most difficult and multi-stage phase of the game. For successful completion, the AI system needed to maintain a single consistent plan over thousands of interrelated steps and more than a day of continuous work, which, according to Google's assessment, is a substantial test for current generative models.
Computational Costs and Resource Intensity
During the experiment, the systems cumulatively used 1.35 billion input tokens. The final computing cost at standard Gemini rates was about $1,620, while at wholesale pricing, the calculation drops to $810. These figures provide a benchmark for the resource intensity of a fully autonomous playthrough of a complex sandbox and can serve as a starting point for future large-scale agent experiments.
Significance of the Result for the Industry
Google specialists emphasize that the success of the experiment demonstrates the ability of modern generative AI models to solve complex problems in complicated, dynamic environments without creating specialized software for each individual task. If the conclusions are confirmed in subsequent independent replications, this could be a turning point for the agentic AI direction: models capable of prolonged self-organization and maintaining a multi-step plan pave the way for automating tasks where formalizing all actions in advance is impossible.
