Google DeepMind announced that a team of four Gemini 4 Argon agents autonomously completed Minecraft by defeating the Ender Dragon, the game's final boss. The experiment, detailed in a paper released by Google DeepMind on October 16, 2026, involved the agents operating in a randomly generated Minecraft survival world on "Easy" difficulty for 27 hours and 14 minutes. During this period, the agents wrote and executed 2,783 programs, exchanged 229 messages, and experienced 605 deaths.
The agents were given a single objective: defeat the Ender Dragon. No human intervention occurred during the playthrough, nor were specific roles, strategies, or game-specific tools assigned in advance. Each agent controlled a player character through Mineflayer, an open-source JavaScript library that allows programmatic interaction with Minecraft. This method meant the agents generated and executed code to navigate the environment, gather resources, craft items, and engage in combat, rather than using traditional keyboard and mouse inputs or visual interpretation.
The Gemini 4 Argon agents demonstrated emergent collaborative behaviors. They quickly began to divide responsibilities, with one agent focusing on tool creation, another on gathering food and supplies, and others on exploration and preparing for more challenging game stages. Communication occurred through in-game chat and a shared memory system, which enabled them to share tips, warn each other of threats, provide coordinates, and even recover items from deceased teammates. For instance, one agent, "janesick," focused on tools, while "gyldenia" gathered food. "Auratian" lit a crucial portal, and "bulicame" constructed Enderman shelters and delivered the final blow to the Ender Dragon.
The research highlights the agents' ability to maintain a coherent plan across thousands of dependent steps over an extended period. The agents also developed 42 reusable helper functions to automate repetitive tasks like movement and resource collection. A fifth agent, designated "the Observer," monitored the session to ensure no cheating or external intervention occurred, and it reported no such incidents.
This achievement builds upon previous AI research in Minecraft, a game often used as a testbed due to its open-ended exploration, resource management, and long-horizon planning requirements. Earlier projects explored reinforcement learning for basic tasks and large language models acting as "brains" for agents that write code. Google DeepMind stated that this is the first known end-to-end autonomous completion of Minecraft by AI agents using only the standard Mineflayer library, without pre-scouted world seeds or custom human-built game tools. The entire process incurred an estimated cost of $810 at Gemini's list price or $1,620 at the standard price, based on the 1.35 billion input tokens used.