The paper was posted on October 15, according to Mashable.
The agents were assigned one goal: kill the dragon. They played on Easy difficulty and completed the run in 27 hours and 14 minutes, with 605 deaths, mostly from monsters. Google says the run used standard tools rather than human-built, game-specific assistance. Researchers monitored a live dashboard and say they did not intervene. Mashable reports an estimated model-use cost of about $520 at Argon's introductory API price.
The result was not a human-equivalent game run. Easy mode reduces monster damage, and the code-based agents could access information such as blocks hidden behind other blocks. The source also says the model's training gave it knowledge of effective Minecraft strategies. It does not establish performance in visually grounded play, harder settings, or real-world agent tasks.