Google says Gemini 4 agents beat Minecraft without human help

The result suggests general-purpose agents can sustain a long task loop without custom game tooling.

Google says four Gemini 4 Argon agents finished Minecraft’s Ender Dragon challenge without human intervention, showing that general-purpose agent systems can now handle a long, complex game loop. The agents ran in a randomized survival world, used only the Mineflayer library, and coordinated through in-game chat and shared memory rather than hand-built Minecraft-specific tooling. Over 27.23 hours they executed 2,783 programs, sent 229 messages, died 605 times, and consumed 1.35 billion input tokens. Google says the run cost $810 at Gemini list price, or $1,620 at standard price, and an independent “Observer” agent watched for cheating and reported on the session. The team called it the first end-to-end autonomous Minecraft completion by AI agents under those constraints.