Google's AI Agents Beat Minecraft's Ender Dragon After 27 Hours

Ever wondered if AI could actually beat Minecraft without any human help? Well, Google just proved it can. Four Google Gemini 4 Argon AI agents have become the first to autonomously defeat the Ender Dragon, taking just over 27 hours to pull it off. That's a world first, according to Google researchers.

But this wasn't a simple speedrun. To even reach The End, the agents had to gather Ender Pearls, craft Eyes of Ender, find a Stronghold, and solve a puzzle. They started with a random seed and only a goal: kill the dragon. No strategy, no roles assigned. Each agent controlled a player using the open-source Mineflayer library, coordinating through in-game chat and shared memory. They planned, wrote, and debugged their own programs.

Surprisingly, the agents quickly fell into roles. Some built Enderman shelters for farming, others gathered food, and some focused on exploration. They shared combat tips and even helped each other recover lost gear when they died. Over the run, they exchanged 229 messages, ran 2,783 programs, and died 605 times. A fifth agent oversaw the process to ensure no cheating, but the high death count suggests they played fair.

It took about 13.5 hours to reach The End, with the remaining time spent trying to kill the final boss (and dying a lot). For comparison, the current Minecraft speedrun record is six minutes and 36 seconds, so these AI agents won't be breaking any records soon. They also used 1.35 billion input tokens, costing around $1,620—or $60 per hour, which breaks down to $15 per hour per agent.

The researchers say this is the first end-to-end autonomous completion of Minecraft by AI agents using only the standard Mineflayer library, without pre-scouted seeds or human-written game-specific tools. "Winning required the agents to keep a coherent plan across thousands of dependent steps for more than a day," they wrote in a paper.

So why do this? The team wanted to study how teams of agents coordinate in a visible setting to understand such systems before trusting them with real-world tasks. The skills needed for Minecraft—teamwork, recovering from failure, and pursuing a goal for over a day—are also what agents need to work reliably alongside people.

What do you think? Are you impressed by these AI agents, or do you think they still have a long way to go? Let us know in the comments!