Four Gemini 4 Argon Agents Beat Minecraft In 27 Hours With No Human Help
The run is a measured cost and time figure for a long-horizon multi-agent task, and the article's own caveats (Easy difficulty, knowledge of Minecraft, 22% of programs ending in errors) define how far that figure can be read.
Reporting from 1 source: GIGAZINE.
Google DeepMind ran a challenge in which four agents on Gemini 4 Argon tried to defeat Minecraft's End Dragon in a randomly generated world, operating through the Mineflayer bot library. They wrote 2,783 JavaScript programs, sent 229 chat messages, died 605 times, and killed the dragon in 27 hours 14 minutes. At standard Gemini 4 Argon pricing, the run cost $1,040 (about 165,000 yen).
The agents split resource gathering and exploration on their own, and used chat to flag dangerous locations and share how they handled failures. When one died, it posted the coordinates and asked the others to recover its items before they vanished. They used triangulation to locate underground fortresses.
The milestone times were uneven: iron tools at 14 minutes, the first Ender Pearl at 3 hours 6 minutes, the first Eye of Ender at 10 hours 43 minutes. The conditions were not human ones. The agents read surrounding blocks and enemies through code rather than a screen, could use information a player cannot see, and played on Easy after learning Minecraft knowledge in advance.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.