They started in a randomly generated survival world on Easy difficulty, with cheat commands disabled, and used the Mineflayer bot library.
The agents wrote 2,783 short JavaScript programs, exchanged 229 in-game chat messages, and died 605 times. They maintained shared notes, divided tasks without assigned roles, and used program outputs such as position, health, inventory, and errors. The researchers estimate 1.35 billion input tokens through the win, costing about $520 at an introductory list price or $1,040 at the standard price.
The result is one run, not a reliability study across tasks or environments. Mineflayer exposed nearby blocks and creatures, including behind solid rock, and a plugin handled route planning. The source also says the researchers could pause or stop the sandboxed run, though they did not need to.