Gemini 4 Argon agents beat Minecraft in 27 hours 14 minutes 12 seconds

The preprint, published on arXiv on October 15, sits against a backdrop of prior Minecraft AI work that either learned pixel-to-keyboard policies (such as VPT, DreamerV3 and SIMA 2) or used language models as planners with human-crafted tools and curricula (such as Voyager and AI Bengineering). The DeepMind team stripped those aids away, leaving four general coding agents to write and share short asynchronous JavaScript programs. The agents maintained persistent state across programs, including a shared memory file and helper functions, and corrected errors through in-game chat.

The most striking structural detail is how the agents handled a navigation failure in the Nether. After one agent proposed a rule forbidding program evaluation at a dangerous coordinate, another agent instead dug out a two-block ledge to make the return path a one-block step, removing the need for the rule itself. The authors frame this as modifying the environment's topology rather than accumulating local prohibitions.

At the same time, the paper itself flags sharp boundaries. The agents read structured telemetry directly from the game's API rather than perceiving pixels, and they relied on prior Minecraft knowledge in their training data. The run also incurred 605 deaths and a long plateau in the End before the team found a minimal-cost tactic of using dirt blocks to climb the crystal towers. The authors explicitly note the result is a single launch they tried, leaving open how often the approach succeeds across random seeds.