Google DeepMind researchers deployed four Gemini 4 Argon AI agents into a randomized Minecraft world with the objective of defeating the Ender Dragon. The agents operated without pre-defined roles, strategies, or external tools, relying solely on the Mineflayer library for bot control and internal communication via in-game chat and shared memory.
During the 27-hour and 14-minute session, the agents autonomously assigned tasks, such as crafting tools and gathering resources, while coordinating survival efforts. They exchanged 229 messages and executed 2,783 programs. Despite suffering 605 deaths, the team successfully defeated the Ender Dragon, with the final blow delivered by the agent named bulicame.
The experiment consumed a total of 1.35 billion tokens. According to Google DeepMind, the cost of the operation reached $810 based on internal Gemini pricing, or $1,620 at standard rates. The research team confirmed that this marks the first instance of AI agents completing the game independently without prior knowledge of world generation seeds or human-made game tools.
The full technical report detailing the performance and behavior of the Gemini 4 Argon agents during this challenge has been published on the arXiv repository. This achievement highlights the capability of large language models to coordinate complex, multi-agent tasks in dynamic environments without human intervention.
