Reproduce “A team of Gemini 4 Argon agents beats Minecraft” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Maxence Faldor, Satinder Singh, Roberta Raileanu
Four Gemini 4 Argon agents completed Minecraft’s survival-mode path to the Ender Dragon in one randomly generated world on Easy difficulty. They wrote programs through Mineflayer, using structured game information and coordinating through chat and shared notes. The run took 27 hours and 14 minutes, with 2,783 programs and 605 deaths; it shows one completion, not reliability across runs.
Key insights
- The agents organized a division of labor themselves, sharing coordinates, hazards, resources, and lessons from failed attempts while adapting their plans.
- Reusable helper functions and stored operational knowledge supported work across thousands of short programs, though errors and repeated deaths remained common.
- The result demonstrates extended coordination through a programmatic interface, not human-like visual gameplay: agents had access to structured state information and a pathfinding plugin.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.