Four Gemini 4 agents beat Minecraft in a 27-hour research run

Four Gemini 4 Argon coding agents defeated Minecraft's Ender Dragon after 27 hours and 14 minutes in a shared survival world, according to a paper submitted October 15. They wrote and ran 2,783 JavaScript programs, exchanged 229 in-game messages and died 605 times before the win. The result is a sustained, recorded example of agents coordinating around one goal. It comes from a single sandboxed run with a programming interface that exposed game state a human player could not see.

The researchers set a narrow objective: defeat the dragon. They did not assign the four players roles or give them a route through the game. The paper says no human intervened in their strategy during the run. BIG CHANGE has not reproduced the experiment or inspected its complete logs.

The big change

• What changed: In an author-reported experiment, four general-purpose coding agents completed a randomly generated Minecraft survival world as a team, from spawn to the Ender Dragon, by writing their own programs through Mineflayer.
• How it happened: The agents read text describing game state, issued JavaScript actions, shared notes and used game chat to divide work and recover from setbacks. A fifth agent observed and reported on the session without playing.
• Why it matters: The run documents coordination across thousands of programs and more than a day of work. Its one-run design, unusually informative interface and repeated failures leave success rate and transfer to other tasks unmeasured.

Four players, one program at a time

The game was Minecraft Java Edition 1.20.4 in survival mode on Easy difficulty, with a random world and cheat commands disabled. Each agent ran on its own cloud machine and controlled a player through Mineflayer and its pathfinding plugin. Rather than see a screen and use a keyboard, an agent wrote a short JavaScript program, received text output about its position, health, inventory, surroundings or an error, then wrote the next program. The game program could control only that character; it could not start processes or write files.

The agents also had a code workspace, shell, web search and persistent private and shared memory on their cloud machines. The authors say none of the 2,839 game programs run over the full session, including after the win, made a network request or referred to a web source. Their general instructions encouraged checking results, building reusable tools and saving lessons. The game-specific instruction was a paragraph giving the goal and allowing the players to decide how to share the work.

The division emerged in chat. One player gathered food and supplies; others worked on tools, a Nether portal and later the stronghold. After an agent drowned carrying twelve Eyes of Ender, teammates relayed the location and the time before the dropped items would disappear; another recovered all twelve. The team later spent nearly 14 hours in the End. It worked out that the ten crystals atop pillars were healing the dragon, destroyed them, then won the fight. These are the authors' account of the recorded session, including server logs, programs and chat.

A successful run with substantial friction

The 27-hour total includes six server restarts that paused the world for about 20 minutes. The paper reports that 22% of the agents' programs ended in an error and that the players died 605 times. A fifth agent, called the Observer, monitored the run and was instructed to intervene only for environmental failures, not strategy. The authors say researchers could pause or stop the game from a dashboard but had no cause to intervene. The Observer and manual review of a sample of logs found no cheating or safety incident, according to the paper.

The paper estimates roughly 24,000 model calls through the win, with about 1.35 billion input tokens, most served from cache, and 16 million output tokens. Applying introductory public API list prices, its authors estimate about $520 for the four players up to the win; they say this is a price-based estimate, not the cost of serving the model. That resource use is part of the setup when judging the result.

What this test leaves open

Mineflayer gave the agents precise information about nearby blocks and creatures, including some hidden behind rock, and even the dragon's internal state. Its pathfinding plugin planned routes. The agents were also drawing on a model trained with substantial prior knowledge of Minecraft and Mineflayer. Easy difficulty reduced the game's penalties. The paper describes the first launch the team tried, with no repeated trials to measure how often this setup would win.

The result therefore supports a specific conclusion: in this environment, this team carried a plan through many dependent steps while adapting to failures. It does not establish a general success rate for long-running agents or show that the same coordination works in software engineering or enterprise operations. Those are analogies the authors draw, and they require separate evidence.