Google DeepMind's Gemini agents beat Minecraft with no human help
A team of four AI agents running on Google's Gemini 4 Argon model has beaten Minecraft by defeating the Ender Dragon, without any help from humans during the game.
The researchers described the game in a research paper published on October 15. The authors were Google DeepMind researchers Maxence Faldor, Satinder Singh and Roberta Raileanu. The four agents started in a randomly created survival world on Easy difficulty. They wrote 2,783 programs and defeated the dragon after 27 hours and 14 minutes. Along the way, they died 605 times.
"To our knowledge, this is the first time AI agents have completed Minecraft from start to finish on their own, using only the standard Mineflayer library, without humans choosing the world in advance or creating special tools for the game," the researchers wrote. Google DeepMind shared the result on X (formerly Twitter) on October 16. The company said the agents "divided up jobs over chat, shared notes, and debugged errors together" without human help.
How the agents played
Each agent controlled a player using Mineflayer, an open-source library. They communicated with their teammates through in-game chat and a shared memory system. According to the research paper, no one assigned them roles or wrote game instructions for them. The agents planned their actions, wrote their own programs and fixed their own mistakes. Researchers watched the game on a live dashboard and could pause or stop it at any time. However, the researchers said they "had no cause to intervene".
The team built a portal to the Nether within the first two hours and found a Nether fortress shortly before the three-hour mark. They had Eyes of Ender after 10 hours and 43 minutes and entered The End after 13 hours and 33 minutes. The final battle took the most time. The agents spent almost another 14 hours fighting before they defeated the dragon.
The agents divided up the work on their own, without being told what to do. One built shelters where they could farm Endermen and delivered the final attack that defeated the dragon. The others collected supplies, managed tools or planned the Nether portal. According to IGN, the four agents exchanged 229 messages in total. A fifth agent watched the game to make sure none of them cheated. The researchers wanted to study how teams of AI agents work together on long tasks. The paper also says that difficult real-world software and business tasks need similar long-term coding and teamwork. It includes a section explaining the limits of the results.