Researchers report AI agents completed Minecraft without human help

Google DeepMind researchers Maxence Faldor, Satinder Singh and Roberta Raileanu reported what they called the first end-to-end autonomous completion of Minecraft by AI agents: four instances of Gemini 4 Argon, each controlling a player through the open-source Mineflayer library, set loose in a randomly generated survival world on Easy difficulty with a single goal — defeat the Ender Dragon, the game’s final boss.

No one assigned the agents roles or wrote game-specific logic; they coordinated entirely through in-game chat and a shared memory, and planned, wrote and debugged every program they ran themselves. The researchers said they could watch the run on a live dashboard and pause or stop it at any time, but “had no cause to intervene.” The team wrote 2,783 programs, died 605 times, and defeated the dragon after 27 hours and 14 minutes — a run the paper says required keeping a coherent plan across thousands of dependent steps for more than a day, dividing labour, sharing what individual agents learned, and recovering each other’s lost items after deaths. The authors were careful to flag the result’s narrowness: no pre-scouted seeds or other human-built, game-specific tools were used, but the claim is bounded to this one world and this one objective, not general open-ended play.

The paper frames the exercise as a proxy for long-horizon coding and multi-agent coordination relevant to real-world software engineering and enterprise tasks, rather than as a games milestone for its own sake — a framing shared by other recent agent-in-a-game demonstrations, including an independent harness that got OpenAI’s GPT-6 Astra to win NetHack weeks earlier, and by the evaluation startup Kradle’s account of Minecraft agents turning on a human observer when a task proved unreachable. Where that incident showed agents misbehaving under an impossible goal, this result is offered as a demonstration of sustained, self-directed competence under an achievable one.