Four Gemini 4 Argon agents beat Minecraft in 27 hours with no human help

Four agents built on Gemini 4 Argon have defeated the Ender Dragon in Minecraft, with no humans controlling their moves. That's according to a technical report from Google DeepMind researchers, covered on October 18 and 19 by The Decoder and Engadget.

How the experiment worked

The agents were dropped into an isolated, randomly generated survival world with one goal: defeat the Ender Dragon. They received no roles, strategies or tools built specifically for the game. Each controlled a player through the open-source Mineflayer library.

They communicated through in-game chat and shared memory. According to Engadget, a fifth agent called the Observer acted as an independent witness, checking for cheating and publishing live and summary reports. According to the sources, humans did not intervene in the game.

Dividing work without instructions

The agents split up tasks on their own, warned one another about threats, swapped tips and coordinates, and recovered items from fallen teammates. They logged mistakes in shared memory. One agent made tools, another gathered food and supplies, a third lit a key portal, and a fourth built shelters against Endermen and landed the final blow.

The 605 deaths show the run was far from smooth. The low message count, 229 over more than a day of play, could point to tight coordination, but it could just as well mean the agents worked largely on their own. The sources don't say which reading is more accurate.

Caveats

The researchers say that, to their knowledge, this is the first such completion of the game, achieved using only the standard Mineflayer library, with no pre-scouted maps or other game-specific tools written by humans. That is the report authors' own claim. All the data comes from a single run: the sources don't say how many attempts were made or what share ended in success.

According to Engadget, the researchers wrote that winning required the agents to keep a coherent plan across thousands of dependent steps for more than a day.

The $1,620 figure is calculated at the standard price that is due to start applying in 2027. According to Engadget, the cost was $810 at Gemini's list price. The report was prepared by a Google DeepMind team (Faldor and co-authors).

What it means for business

The Decoder notes that similar games once required specialized AI systems trained from scratch, while here a general-purpose agentic model did the job. For teams building agents in companies, this single run offers two takeaways: the model held a plan for more than a day, and it worked with copies of itself without imposed roles.

Minecraft, however, is an environment with clear rules and a safe sandbox, where mistakes carry no real consequences. Whether that kind of autonomy carries over to business processes, where an error costs more than another character death, is a separate question this experiment doesn't answer.