Four Gemini 4 Argon Agents Teamed Up for 27 Hours to Kill a Minecraft Boss

Google DeepMind researchers have published an experiment in which four AI agents running Gemini 4 Argon defeated the Ender Dragon, Minecraft’s final boss. According to the paper released on October 15, the agents worked together for 27 hours and 14 minutes without any human intervention. The researchers gave them one goal: defeat the dragon. They did not assign roles or lay out an in-game strategy in advance.

How the four agents worked together

The four agents, which handle tasks on their own, exchanged information through in-game chat and a shared memory. They divided up the roles they needed among themselves, retrieved items that teammates had lost, and adjusted combat tactics as they pushed toward the goal. In total they executed 2,783 programs and sent 229 chat messages.

What drove the game was code, not a screen and keyboard. The agents wrote and ran programs using Mineflayer, a tool for controlling the game. Each action’s result came back as text, and the agents used that to decide what to do next. They were also given a pathfinding tool for navigation. The researchers say they confirmed the dragon kill through the achievement recorded in the game server logs.

605 deaths: how reliable is the result?

Getting there took a lot of trial and error. In a randomly generated world on easy difficulty, the characters died 605 times in total. Of the programs the agents executed, 22% ended in errors, and the server had to be restarted six times. The paper reports a single run.

Source: Google (press release)

The approach here differed from earlier research. Previous efforts trained models on gameplay data, or equipped them with game-specific planning tools and skill libraries. This time, general-purpose coding agents were given only basic game-control tooling, a pathfinding tool, and a brief description of the goal, and then left to collaborate.

What the agents could see was not what an ordinary player sees. They read information through code that a human cannot see on screen, such as block positions or the dragon’s internal state. The paper also notes dragon-kill reports from other projects.

What did it cost?

The researchers estimate that the usage required to kill the dragon would come to roughly $520 at public list prices. The full session, including an observer agent and post-victory activity, works out to about $810.

Gemini 4 Argon was formally released on October 13. It is available through the Gemini API, which developers plug into their own services, and through Google AI Studio, and both services’ supported regions include South Korea. It is not available on the free tier. Pricing runs at $2 per million input tokens and $10 per million output tokens through November 30. From December 1, those rates rise to $4 and $20 respectively.

If you are thinking about putting collaborative AI to work, it is worth looking at how the agents divided up roles alongside the completion time, error rate, and cost. The open questions are whether the same result can be reproduced under the same conditions, and whether this kind of collaboration holds up on tasks outside a game.