[Google’s Gemini 4 Agents Defeat Minecraft Without Human Intervention]
Agents split up tasks, warned about threats and even recovered items from dead teammates.
<
Google’s DeepMind “AlphaGo” AI is famous for winning at Go, but the company’s latest Gemini AI just mastered a far more complex game. Four Gemini 4 Argon agents beat Minecraft’s final “Ender Dragon” foe with no human intervention by working individually while sharing tips, threats and mistakes, Google researchers revealed in a new paper. The experiment shows how general agentic AI models can now beat complex games, a feat that used to require custom AI software.
The team deployed four Gemini 4 Argon agents within a sandboxed, randomly generated Minecraft survival environment, whose sole objective was to slay the Ender Dragon. Assigned no predefined roles, strategies, or bespoke game utilities, the agents each managed a Mineflayer‑controlled avatar and communicated via in‑game chat plus a shared memory store.
Throughout play the agents soon divided responsibilities—alerting one another to potential threats, offering guidance and lessons learned, supplying critical coordinates, and retrieving belongings left behind by fallen companions. The team naturally emerged into distinct functions: janesick specialized in tool creation, gyldenia focused on gathering food and resources, auratian manned the pivotal portal, and bulicame constructed Enderman shelters before delivering the decisive strike.
From start to finish the collaboration lasted 27.23 hours, the bulk of which was devoted to confronting the Ender Dragon. The group contributed 2,783 executed programs, sent 229 messages, and perished 605 times during the trial. To oversee fairness and generate live summaries, a fifth agent named “the Observer” acted as an impartial auditer. The agents consumed 1.35 billion input tokens, and the project cost $810 at Gemini’s list price or $1,620 at the base rate.
“To our knowledge, this represents the first fully autonomous Minecraft conquest by AI agents, accomplished solely with the standard Mineflayer implementation and without any pre‑scouted seeds or game‑specific human‑authored tooling,” the research team explained. “Securing the victory demanded that the agents maintain a coherent plan across thousands of interdependent steps over more than a day.”
