Four Gemini Agents Beat Minecraft Alone — 27 Hours, 605 Deaths, $810

Four Google Gemini 4 Argon agents defeated Minecraft's Ender Dragon in 27.23 hours, with no human intervention, at a list-price cost of $810. The run, described in a new paper by Google researchers, marks what the team calls the first end-to-end autonomous completion of Minecraft by AI agents using only the standard open-source Mineflayer library — no pre-scouted seeds, no hand-written game-specific tools.

The result matters beyond gaming. Minecraft has long served as a benchmark for AI research because it demands long-horizon planning, exploration and resource management. Beating it used to require custom AI software. A general-purpose model doing it with off-the-shelf tooling signals that agentic AI has crossed a threshold in sustained, coordinated task completion.

"To our knowledge, this is the first end-to-end autonomous completion of Minecraft by AI agents, achieved using only the standard Mineflayer library without pre-scouted seeds or other game-specific tools written by humans," the team wrote. "Winning required the agents to keep a coherent plan across thousands of dependent steps for more than a day."

How did the agents pull it off?

Google's DeepMind built its reputation on AlphaGo, which beat world champions at Go. Minecraft is a messier challenge — open-ended, randomized and far longer than a board game.

The researchers dropped four agents into a sandboxed, randomized survival world with a single goal: kill the Ender Dragon. They assigned no roles and no strategies. Each agent controlled a player through the Mineflayer library and coordinated with teammates via in-game chat and shared memory.

Left to themselves, the agents organized. They split tasks, warned each other about threats, traded tips and coordinates, and recovered items from fallen teammates. The division of labor that emerged:

• "janesick" made tools
• "gyldenia" gathered food and supplies
• "auratian" lit a key portal
• "bulicame" built Enderman shelters and landed the final blow

The numbers behind the run are stark. "The four agents wrote and ran 2,783 programs, exchanged 229 messages and died 605 times along the way," the team said. Most of the 27.23 hours went to the Ender Dragon fight itself. Total compute: 1.35 billion input tokens.

Was anyone watching for cheating?

Yes. A fifth agent, called "the Observer," monitored the session as an independent witness. Its job was twofold: check for cheating and publish live and summary reports. That safeguard matters as autonomous agents take on longer, harder tasks where humans can't supervise every step.

The cost figures are notable too. The run cost $810 at Gemini's list price, or $1,620 at the standard price. That puts a full autonomous Minecraft completion within reach of a modest research budget — a data point for anyone pricing out long-running agent deployments.

Why does this matter for AI beyond games?

The core achievement is coherence. Keeping a plan intact across thousands of dependent steps for more than a day is exactly the failure mode that has plagued agents in real-world settings, where tasks fragment and context breaks down.

The Minecraft run shows general models can now do what previously demanded bespoke software: divide labor, communicate, adapt to teammates' deaths and finish the job. That has obvious implications for multi-agent systems in software development, logistics and research automation, where the same coordination problems appear without the pixelated dragons.

The next question is transfer. Google's paper demonstrates the capability in a game world with clean rules and a shared chat channel; whether the same four-agent architecture holds up in messier real-world environments is what researchers and competitors will now test.