The Pulse
Four Gemini 4 Argon Agents Beat Minecraft’s Ender Dragon
Four Gemini 4 Argon agents defeated Minecraft’s Ender Dragon in a randomly generated Easy-mode world after 27 hours and 14 minutes. The researchers report 2,783 programs, 605 deaths and a single completed run using Mineflayer.
AI.info Team · Written with AI ·
The four Gemini 4 Argon agents died 605 times before defeating Minecraft’s Ender Dragon. The run took 27 hours and 14 minutes, involved 2,783 programs written by the agents and ended in a randomly generated world on Easy difficulty, according to a paper submitted to arXiv on October 15.
The researchers describe the result as, to their knowledge, the first end-to-end autonomous completion of Minecraft by AI agents using the standard Mineflayer library, without a pre-scouted seed or human-written game-specific tools. It was one run, not a measure of how reliably the system can finish the game.
Four players, one shared goal
Each agent controlled a player through Mineflayer, an open-source JavaScript library. The agents exchanged information through in-game chat and shared memory, and decided how to divide the work themselves. Researchers watched through a live dashboard and could pause or stop the session, but say they did not need to intervene.
“Starting from a randomly generated world on Easy difficulty, the team wrote 2,783 programs and defeated the dragon in 27 hours and 14 minutes, dying 605 times along the way.”
A fifth Gemini 4 Argon instance served as an Observer rather than a player. The paper says all programs, chat messages and server events were recorded; the win also appears in the server log as the “Free the End” advancement.
Mineflayer gave the agents more than a view of the game
The players did not control Minecraft through screenshots and button presses. They wrote JavaScript programs that Mineflayer ran, then read text describing their position, health, inventory, surroundings and program results. Its pathfinding plugin planned routes, while the library exposed information such as nearby blocks and creatures.
That interface shaped the result. The authors note that Mineflayer could reveal blocks behind solid rock and the dragon’s internal state—information unavailable to a human player looking at the screen. The agents also had prior knowledge of Minecraft and Mineflayer from training, and the game ran on Easy, where monsters deal less damage and hunger cannot kill a player.
The team turned setbacks into shared instructions
The agents split tasks across the Overworld, Nether and End. One located a Nether fortress and later calculated the stronghold’s position by tracking two thrown Eyes of Ender; the estimated location was 43 metres from the portal room. Teammates shared coordinates and recovered dropped items after deaths.
The dragon fight exposed a different problem: its health kept recovering. About 21 hours into the run, an agent identified the ten crystals on obsidian pillars as the source of the healing. The team stacked dirt beside the pillars, punched the crystals and dug back down; all ten were destroyed over the next four hours and 40 minutes. The dragon fell 39 minutes after the last crystal was removed.
A win, not a reliability test
The paper records 229 exchanged messages and six server restarts, which paused the world for about 20 minutes in total. It also reports roughly 24,000 model calls and estimates the four players’ run cost at about $520 using Gemini 4 Argon’s introductory public API list price, or $1,040 at the standard price. The full session, including the Observer and activity after the win, is estimated at $810 or $1,620, respectively; the authors stress that these are list-price estimates, not the cost of serving the model.
The study tests whether general-purpose coding agents can turn existing game knowledge into programs and coordinate through a long task—not whether they can learn Minecraft from scratch. Because the researchers describe only their first launch, the result cannot show how often another team would succeed. It shows that one four-agent team completed one Easy-mode run, with the server recording the final “Free the End” advancement 27 hours and 14 minutes after its first program.