Google DeepMind has published a paper about their AI team autonomously finishing the game Minecraft. The team is the first to end-to-end play the game through to the end with AI agents. No human game logic was written.
Four Gemini 4 Argon agents coordinate roles and write code to defeat Ender Dragon
No pre-scouted seeds were done. This is just an experiment that shows how frontier models can deal with complex and open-ended worlds. This shines a light on the limits of autonomous agents. They used a team of 4 agents running on Gemini 4 Argon model to complete the task.
They used the Mineflayer JavaScript library to allow the AI agents to engage with the game. The team had different agents take on certain roles to optimize their effort. They had an agent named Gyldenia in charge of food, beds, and flint.
They had an agent named Bulicame focused on diamonds. Janesick was tasked with iron and tool crafting. Finally, they had an agent named Auratian who was in charge of setting up the Nether portal. So they had the team split up to divide the work and step through the various phases of the game systematically.
Overall, the team played through the game for 27 hours and 14 minutes. This was from program start to final boss kill. The team ended up dying 605 times while playing on Easy mode.
They produced 2,783 total programs to help solve in-game problems. This ended up using about 1.35 billion input tokens and 16 million output tokens. So overall, they made about 24,000 calls to the model. So on promotional pricing, this ended up costing $520 for the tokens.
They ran into some technical challenges to beat the Ender Dragon. They found a trick with using dirt pillars to get to the dragon's healing crystals. This allowed them to break the crystals and take down the boss.
Google DeepMind said the game didn't stand a chance to their team of 4 agents. They claim this is the first autonomous completion of Minecraft in this way. There was an extra observer agent that was used during the reporting phase but not for playing.
So if you factor in the observer agent, then the total session cost goes up to $810. That would increase to $1,620 at list prices. So this is an interesting experiment to show that AI agents can team up and take different roles.
They can also figure things out in the sandbox game world. So it shows the power of the Gemini 4 Argon model to do long horizon tasks.
Discussion
0 comments
Log in to join the thread with a thoughtful take, question, or correction.