Google DeepMind Agents Beat Minecraft’s Ender Dragon Autonomously in 27 Hours
Four general-purpose Gemini 4 Argon coding agents, coordinating only through in-game chat and shared memory, defeated the Ender Dragon from a random Easy-difficulty spawn in 27 hours 14 minutes, writing 2,783 programs—authors say the first end-to-end autonomous Minecraft completion using only standard Mineflayer.
By 1EyeAI Desk ·
Google DeepMind researchers Maxence Faldor, Satinder Singh and Roberta Raileanu describe in an arXiv preprint (abs posted mid-October 2026) a run in which four Gemini 4 Argon general-purpose coding agents shared one Minecraft Java 1.20.4 survival world on Easy difficulty. Each drove a player through the open-source Mineflayer library and its pathfinding plugin, talking only via in-game chat and a shared memory. Roles were not assigned; no human-written game tools or pre-scouted seed were added. Researchers could pause the run but report they did not intervene.
From a randomly generated world, the team wrote 2,783 JavaScript programs, exchanged 229 messages and died 605 times before defeating the Ender Dragon in 27 hours and 14 minutes. The agents specialized (scouting, tools, Nether, End crystals), recovered teammates’ items after deaths, and posted rules after failures. A fifth Observer agent watched for cheating and reported none. Token use up to the win is estimated at about $520 at Gemini 4 Argon’s introductory list price ($1,040 at standard list)—authors’ estimate, not a bill.
The authors claim this is, to their knowledge, the first end-to-end autonomous completion of Minecraft by AI agents using only standard Mineflayer without human-written game-specific tools. They also note community projects that reported dragon kills with different setups (high-level tool packs or fixed seeds). Limitations they list: Mineflayer gives privileged text state (including some hidden game internals), Easy difficulty, prior Minecraft knowledge in the model, many failures, six server restarts, and a single run.
What this means for you: Nothing ships in a consumer product from this paper alone. If you build coding agents, the takeaway is that multi-agent coordination, shared memory and recovery loops can matter as much as the base model—and that sandbox demos still need careful reading of what the agent can “see.”