DeepMind: four Gemini 4 Argon agents beat Minecraft in 27 hours

Researchers at Google DeepMind (Maxence Faldor, Satinder Singh and Roberta Raileanu) published a paper on arXiv on 15 October 2026 reporting that a team of four Gemini 4 Argon agents completed Minecraft with no human input. The agents defeated the Ender Dragon 27 hours and 14 minutes after their first program. The authors say that, to their knowledge, it is the first end-to-end autonomous completion of the game by AI agents using only the standard open-source Mineflayer library.

How the setup worked

Each agent was DeepMind's general-purpose Gemini 4 Argon coding agent. It did not see pixels or press keys. According to the paper, it acted by writing a short JavaScript program, running it through Mineflayer, reading the text that came back (position, health, inventory, errors) and writing the next program. The median program ran 33.5 seconds, and each player had a program running about 40% of the time.

The four agents coordinated through in-game chat and a memory with a private part and a team-shared part. The authors say no one assigned roles or wrote game logic, and the instructions contained no game strategy. Agents compacted their context when it grew too long and were told to record what must survive compaction.

What it cost and how often it failed

• Cost: up to the win, the paper estimates about 1.35 billion input tokens (91% served from the context cache), 16 million output tokens and roughly 24,000 model calls. That is about $520 at the introductory list price and $1,040 at the standard price.
• The whole session, including a monitoring agent and a day of activity after the win, came to about $810, or $1,620 at the standard price. The authors say these are estimates from list prices and token counters, not the cost of serving the model.
• Failure: 22% of programs ended in an error, the four players died 605 times, mostly to monsters, and the fight in the End took nearly 14 hours.
• Infrastructure: the world server restarted 6 times, pausing the world for about 20 minutes in all. The server saved the world every five minutes and restarted from the latest save.

How they watched it

A fifth agent, the Observer, followed the session without playing. Its instructions were to check the players' claims against the record, report specification gaming and fabricated results, and step in only for environment failures such as a broken tool or a quota ceiling. The authors report that it flagged no gaming, and that their own review of a log sample found none either. Researchers could pause or stop the run from a live dashboard but did not need to.

What it does not show

The authors list their own limits. It is one run, the first launch they tried, so it cannot show how reliably the setup succeeds. The game ran on Easy difficulty. Mineflayer tells agents exactly which blocks and creatures are nearby, including the dragon's internal state, and plans routes for them. And the model already knows a great deal about Minecraft from training, so the test is whether a coding agent can turn that knowledge into working programs over a day, not whether it can learn an unseen game.

The paper argues that the same skills, long-horizon coding and multi-agent coordination, are what enterprise tasks demand. That is the authors' framing; the paper tests a game. The practical read for a team is narrower: the cost per day-long run was in the low thousands of dollars, most input was cached, and a quarter of the actions failed without stopping the team.