Four Gemini 4 Argon agents reportedly defeated Minecraft’s Ender Dragon in a single run lasting 27 hours and 14 minutes, with 2,783 programs written and 605 deaths recorded. A paper submitted to arXiv on October 15, 2026 describes the run. Researchers could monitor it and pause or stop it, and the paper says they had no cause to intervene.
How the agents played and coordinated
The paper describes a randomly generated Minecraft Survival world on Easy difficulty. The four agents used Mineflayer, an open-source library for interacting with Minecraft, and coordinated through in-game chat and shared memory. They planned, wrote, and debugged their own programs; researchers did not assign roles or write game logic.
APPSO, in an article republished by 36Kr, reported that the experiment used Java Edition 1.20.4 and JavaScript. That account also described about 24,000 model calls, 1.35 billion input tokens, and 16 million output tokens. It reported that about 22% of the programs contained errors and that the server restarted six times.
Researchers monitored the run and could stop it
The paper says researchers watched through a live dashboard and could pause or stop the run, but did not intervene. The agents therefore completed the reported task without researchers stepping in to direct play, under human monitoring with a stop option.
One win does not establish reliable completion
The paper describes one successful run. APPSO’s account, republished by 36Kr, says repeated tests on random maps were insufficient to establish stable success. The reported result does not establish how often the agents could finish other randomly generated worlds.
The cost estimates use different pricing assumptions
APPSO/36Kr estimated the run at about $520 using discounted API pricing. IGN separately estimated that the reported 1.35 billion input tokens would cost about $1,620 at normal-user pricing. The estimates use different pricing assumptions, and neither is a verified invoice.
