Gemini 4 Argon Agents Defeat Minecraft’s Ender Dragon in 27 Hours and 14 Minutes

Four Gemini 4 Argon agents defeated Minecraft’s Ender Dragon in 27 hours and 14 minutes, completing one long survival run through shared planning, code and repeated recovery from failure.

The agents played Minecraft Java Edition 1.20.4 in survival mode on Easy difficulty, in a randomly generated world with cheats disabled. They acted through code, with no human input during play. Researchers monitored the session and could pause or stop it, but did not intervene.

Each agent controlled a player through Mineflayer, a JavaScript library. Instead of interpreting screenshots, the agents received structured information about their characters and the game world, including blocks behind solid rock and the dragon’s internal state. A pathfinding plugin planned routes. The agents communicated through in-game chat and shared notes, without assigned roles.

To reach the dragon, the team gathered and crafted supplies in the Overworld, collected resources in the Nether and found a stronghold leading to the End. The agents identified 10 crystals restoring the dragon’s health, then stacked dirt blocks to reach and destroy them. They wrote and ran 2,783 programs, exchanged 229 chat messages and died 605 times. The game server restarted six times, pausing the world for about 20 minutes in total.

The agents produced 63,313 distinct lines of JavaScript and developed 42 helper functions. One agent estimated the stronghold’s position to within 43 meters of the portal room. About 22% of programs ended in an error, showing how much recovery the run involved.

Through the victory, the agents made roughly 24,000 model calls, processing about 1.35 billion input tokens and generating 16 million output tokens, including 4 million thinking tokens. About 91% of input tokens were served from cache. Applying API list prices to those token totals gives estimated costs of about $520 at introductory rates or $1,040 at standard rates. Those estimates do not measure the cost of serving the model.

The experiment was a single attempt. It does not establish how reliably the agents could repeat the run or whether the approach would work on real-world software tasks; they had detailed access to game data and played on Easy difficulty.