Gemini 4 Argon agents beat Minecraft in 27 hours

The dragon is the headline. The 605 deaths are the lesson.

Four AI agents running on Google's Gemini 4 Argon finished Minecraft, killing the Ender Dragon, without human help, according to a paper posted to arXiv on Oct. 15 by Maxence Faldor, Satinder Singh and Roberta Raileanu. The Decoder, which covered it on Sunday, attributes the work to Google DeepMind.

How the run worked

The four players started in a randomly generated survival world on Easy difficulty, with cheats off, the paper says. A fifth Gemini agent watched the run and never played.

Each player controlled a character by writing and running JavaScript through Mineflayer, an open-source library for Minecraft bots, with its pathfinding plugin. The authors say they gave the agents no roles, no curriculum, no skill library and no game-specific tools written by people. The agents coordinated through in-game chat and a shared memory.

Researchers could watch a live dashboard and pause or stop the run. The paper says they "had no cause to intervene."

The time to beat the game:

27 hours and 14 minutes.

Along the way, the paper counts 2,783 programs, 229 chat messages and 605 deaths, 69 of them in the End. About 22% of programs ended in an error. The final fight took nearly 14 hours.

The authors estimate the player agents cost about $520 at Gemini 4 Argon's introductory price, or $1,040 at standard pricing. The whole session, including the observer and activity after the win, came to about $810, or $1,620 at standard prices. Those are list prices, the paper notes, not serving costs.

Most of the input was reused. The players read about 1.35 billion input tokens, 91% of them cached, and wrote about 16 million.

Stamina is the capability

Yes, it's a video game. The skill it tests is the one your long agent jobs need.

Nobody should care that an AI can kill a dragon. You should care that four agents kept a plan coherent across thousands of dependent steps, divided the work themselves, and kept going after hundreds of failures, for more than a day. The authors make the same link to long-horizon coding and multi-agent software work.

The cost line is the practical part. With prompt caching carrying 91% of the input, a day of four agents came to about a thousand dollars at standard prices. Google's introductory pricing, which TopFive covered last week, halves it.

Read the limits before you read the claim

The authors list their own caveats, and they are real.

Mineflayer hands the agents exact information about blocks and creatures, including blocks hidden behind rock. Playing from pixels with a keyboard and mouse "remains much harder," the paper says. Easy difficulty means less damage, and hunger can't kill. The agents' Minecraft knowledge comes from the model's training data. And this was one run, the first they launched, so it says nothing about how often the setup succeeds.

It is also not the first dragon kill by an AI. The paper cites a single Claude Opus 5 agent that did it on Normal difficulty, and a planner paired with an action model on a fixed speedrun seed on Peaceful. The authors claim, "to our knowledge," the first end-to-end autonomous completion of Minecraft by AI agents without pre-scouted seeds or game-specific tools written by people.

What to watch

Watch for a repeat: more runs, harder difficulty, other models in the same harness. One success is a demo. A success rate is a capability.

Then price your own long jobs per finished task, retries included.