Google DeepMind’s Gemini 4 Argon AI agents beat Minecraft in 27 hours

Google DeepMind has shown that general-purpose artificial intelligence (AI) agents can handle difficult tasks that take a long time. Four agents powered by its Gemini 4 Argon model completed Minecraft in 27 hours without human help.

The agents began in a survival world created at random and defeated the game's final boss, the Ender Dragon, after creating 2,783 programs to complete the challenge. They worked in a controlled computer environment, using code and text messages to plan together, share information and fix mistakes.

According to Google DeepMind, the experiment did not use game instructions written in advance or human help. People only watched to make sure the test was safe. The agents worked out for themselves what code they needed, wrote it and ran it to move forward in the game.

“Four Gemini 4 Argon agents beat Minecraft from scratch in 27 hours,” Google DeepMind said. The company explained that the agents divided the work through chat, shared notes and fixed programming errors together.

Testing AI agents beyond systems made for specific games

This achievement is different from DeepMind's earlier gaming successes, including AlphaGo and AlphaStar. Those systems were trained specifically for individual games. In this experiment, the company used general-purpose coding agents like the ones engineers use for everyday programming tasks.

Minecraft gave the agents many connected challenges that required them to follow a clear plan for a long time. They had to collect and prepare materials, find rare resources in dangerous alternate dimensions, use calculations about space and distance to locate an underground fortress, and finally defeat a flying enemy that could restore its health.

Making mistakes could also cost them time. If a player died, they could lose the items they were carrying. This might mean spending hours getting those items back before continuing.

DeepMind designed the test to find out whether ordinary coding models could complete every stage without taking shortcuts. They could not use maps explored in advance, portals opened beforehand or scripts with game instructions already written into them.

What this could mean for complex tasks in the real world

The company said the experiment provides evidence that independent AI systems can work together, recover from repeated failures and keep working toward goals that require many steps to depend on one another.

According to DeepMind, these abilities could be useful for tasks beyond gaming, including software engineering, scientific research, robotics and complex physical systems.

The final battle showed that the agents could keep going even under pressure. Thirty-nine minutes after destroying the last crystal that healed the Ender Dragon, the team defeated it. The agent that landed the final hit had only 2.5 health points left out of a maximum of 20.

Follow Storyboard18 on Google to get the latest and breaking digital news and industry updates, as well as detailed coverage of digital media and gaming news. Get the latest views and updates from Storyboard18.