Google AI completes Minecraft in 27 hours using four autonomous agents

A team of four AI agents, powered by Google’s Gemini 4 Argon model, defeated Minecraft’s final boss, the Ender Dragon, after 27 hours and 14 minutes of gameplay. During the session, the agents recorded 605 deaths, created 2,783 JavaScript programs, and exchanged 229 messages via the game’s chat before securing victory.

How the agents operated

The agents controlled their characters using Mineflayer, a programming library that provides information about the virtual environment and enables actions to be executed via code. Each agent developed its own scripts without receiving a predefined strategy from researchers. Coordination among them emerged organically: while some gathered resources, others warned of dangers and retrieved items left behind by teammates who had died. All lessons were recorded in a shared memory, which was used to guide future attempts.

Google DeepMind described the experiment in an article published on Thursday, the 15th. Researchers could monitor or interrupt the process, but the agents acted autonomously, selecting tools, executing actions, and evaluating results before deciding their next move. The model was already familiar with Minecraft, having been trained on game data.

What does it mean to “complete” Minecraft

In Minecraft, “completing” the game is commonly understood as defeating the Ender Dragon. After victory, the world remains available for exploration and building, and the player can access the final credits.

Limitations of the result

• The test was conducted in Easy mode, with Mineflayer providing data that a human player would not have access to.
• This was a single execution, which does not allow for estimating how often the team would repeat the feat.
• The setup differs from other projects that have defeated the dragon using AI, as it employs distinct conditions.

Therefore, the result does not equate to playing Minecraft in the same way a human would, by looking at the screen and using a keyboard and mouse.

Conclusion

The study demonstrates the potential of AI agents to perform complex and coordinated tasks in virtual environments. At the same time, it highlights current limitations, such as the reliance on controlled conditions and additional data provided by specific libraries.