Four Gemini 4 Argon agents beat Minecraft in 27 hours without human help

Four Gemini 4 Argon agents defeated the Ender Dragon in Minecraft in 27 hours and 14 minutes, starting from a randomly generated world on Easy difficulty, according to a paper posted to arXiv on 15 October 2026 by Maxence Faldor, Satinder Singh and Roberta Raileanu [1].

How did the Gemini agents play Minecraft?

Each agent controlled one player through Mineflayer, an open-source library, and coordinated with teammates through in-game chat and shared memory [1]. No one assigned roles or wrote game logic, and the agents planned, wrote and debugged every program themselves [1]. Researchers could pause or stop the run from a live dashboard but did not intervene, the authors write [1].

Along the way the four agents wrote 2,783 programs, exchanged 229 messages and died 605 times; the run included 6 server restarts and spanned about 81 in-game days, the paper says [2].

Why does a Minecraft run matter for AI agents?

The authors say winning required keeping a coherent plan across thousands of dependent steps for more than a day, and that complex software engineering and enterprise tasks demand the same long-horizon coding and multi-agent coordination [1]. The team divided work, shared what it learned and recovered each other's items [1].

Our read

Opinion: Minecraft is the demo, but stamina is the result. Most agent benchmarks still end in minutes, and this run shows where testing needs to go: plans that survive hundreds of failures over a full day.

What we could not verify

• The paper describes a single run on Easy difficulty; we have not seen replications.
• The paper does not say how much compute or money the run cost, as far as we read.