Four Gemini 4 Argon agents beat Minecraft on their own

Researchers dropped four Gemini 4 Argon agents into a randomly generated Minecraft survival world on Easy. They gave them one goal, killing the Ender Dragon, and no roles, strategy or game-specific tools. The paper's authors are Maxence Faldor, Satinder Singh and Roberta Raileanu, and The Decoder credits Google DeepMind. The agents played by writing code with the open-source Mineflayer library and coordinated through in-game chat and shared notes. They won in 27 hours and 14 minutes, after writing 2,783 programs, sending 229 messages and dying 605 times. The researchers could pause the run at any point but say they never had to step in. The paper calls it the first fully autonomous Minecraft completion by AI agents.

What this means for you

The paper estimates the whole session cost about $810 at Argon's current introductory API price, or about $1,620 at the standard price. Those are list-price estimates, not what it costs to serve the model. One caveat is that Mineflayer hands the agents exact block and mob data, even through solid rock, so they weren't playing from the screen the way a person does. The real takeaway is that the agents kept a long-running team coding effort on track for more than a day, and the authors say that's what matters for software work.

Sources: arXiv paper and The Decoder