A new Google DeepMind study takes general-purpose AI agent teamwork to a new level: four agents built on Gemini 4 Argon completed Minecraft from scratch in 27 hours and 14 minutes with zero human intervention, finally slaying the Ender Dragon. The paper appeared on arXiv on October 15 and was picked up by tech media starting October 20.
Point 1: No intervention, self-organized division of labor. The agents were dropped into a randomly generated survival world with a single goal — defeat the Ender Dragon. The researchers assigned no roles, gave no strategy, and wrote no game-specific tools. The agents drove their characters through the open-source Mineflayer library and coordinated via in-game chat and shared memory: they divided the work on their own (crafting tools, gathering food, building the portal, fighting the boss), warned each other of threats, shared coordinates and tips, recovered items dropped by fallen teammates, and logged mistakes into shared notes.
Point 2: An independent "Observer" supervised everything. A fifth agent served as an independent witness, monitoring the whole run for cheating or rule-bending while publishing live reports. Verdict: zero violations.
By the numbers: 2,783 programs, about 63,000 lines of code, 229 messages, 605 deaths. The team calls it the first end-to-end autonomous Minecraft completion by general-purpose AI agents — earlier feats of this kind relied on reinforcement-learning models custom-built for the game. Keeping a coherent plan across thousands of interdependent steps for more than a full day is exactly the long-horizon multi-agent coordination that complex software engineering and enterprise tasks demand.
Source: Read the original. This is a summarized digest for reference only.