Google DeepMind just dropped four Gemini 4 Argon agents into a random Minecraft survival world. No assigned roles. No strategies. No game-specific tools beyond the standard Mineflayer library. Then they watched.
27 hours and 14 minutes later, the Ender Dragon was dead.
Along the way the four agents wrote 2,783 JavaScript programs, swapped 229 chat messages, died 605 times, and organized themselves into emergent roles (one even handled cleanup duty). A fifth "Observer" agent audited the run. Total input: 1.35 billion tokens. Cost at Argon's standard pricing: about $1,620 for the full session.
Why it matters: nobody trained these models to play Minecraft. They reasoned through crafting recipes, split mining and building duties, warned each other about threats, and recovered gear from fallen teammates, all through text and code. That loop — plan, code, fail, revise — is exactly what agentic AI is supposed to do in software development and robotics, not just games.
The honest caveats: it was one run on easy difficulty, Argon already "knows" Minecraft from training data, and a repeat rate is unknown.
Full breakdown, cost math and FAQs: https://toolpilotpro.com/gemini-4-argon-agents-beat-minecraft
Top comments (0)