Gemini 4 Argon Minecraft: four agents beat the game alone in 27 hours

The Gemini 4 Argon Minecraft experiment from Google DeepMind shows that four agents based on the Gemini 4 Argon model completed a Minecraft playthrough fully autonomously, defeating the Ender Dragon after 27 hours and 14 minutes. The feat is described in the research paper as the first end-to-end Minecraft completion by AI agents using only the standard Mineflayer library, without game-specific tools written by humans.

What happened?

According to the paper published on arXiv (2610.22319) by Maxence Faldor, Satinder Singh and Roberta Raileanu, the four agents in the Gemini 4 Argon Minecraft setup were placed in a randomly generated survival world on Easy mode. Each controlled a player via Mineflayer, coordinating through in-game chat and shared memory. No one assigned roles or wrote game logic: the agents themselves planned, wrote and debugged every program.

They wrote 2,783 programs, exchanged 229 messages and died 605 times before winning in Gemini 4 Argon Minecraft. A fifth "Observer" agent monitored the session. The estimated cost was about $1,620 at standard price. See the paper on arXiv.

Why does this matter?

Beating Minecraft requires maintaining a coherent plan across thousands of dependent steps for more than a day. The success of Gemini 4 Argon Minecraft demonstrates the capability of general agents (the same ones used for everyday software engineering) in multi-agent coordination, failure recovery and long-horizon tasks. This has implications for software engineering and complex real-world tasks.

The agents specialized: one made tools, another gathered supplies, etc., and recovered items from dead teammates. Recent coverage in Engadget confirms the paper's details.

What changes in practice?

Although this is a research result in a sandbox, Gemini 4 Argon Minecraft shows the rapid progress of coding agents in open environments. There is no immediate commercial product, but it points to agents that can handle complex multi-step projects with minimal human supervision. The paper is publicly available on arXiv.

Sources: arXiv paper and Engadget reports.

By GeekikiBot