About the news
Researchers Maxence Faldor, Satinder Singh, and Roberta Raileanu placed the agents in a sandboxed Minecraft Java Edition world on Easy difficulty. Cheat commands were disabled. The agents received one clear goal: defeat the Ender Dragon. No roles were assigned, no strategy was provided, and no game-specific tools were written by humans in advance.
Each agent controlled a player through Mineflayer, an open-source JavaScript library for Minecraft bots. The agents never saw the game screen or pressed keys. Instead, they acted by writing short programs, running them, and reading the text results that came back — position, health, inventory, or error messages. They coordinated with each other through in-game chat and shared memory.
Key results from the run:
- Time to defeat the Ender Dragon: 27 hours and 14 minutes
- Programs written: 2,783
- Player deaths: 605
- Chat messages exchanged: 229
- The agents spontaneously divided work (tools, supplies, Nether portal, stronghold search, crystal destruction, and the final fight)
- They recovered items after teammate deaths and recorded lessons between attempts
- Researchers monitored a live dashboard and could pause or stop the run, but did not need to intervene
Why it matters
The authors state that, to their knowledge, this is the first end-to-end autonomous completion of Minecraft by AI agents using only the standard Mineflayer library, without pre-scouted seeds or other human-written game tools. Winning required the agents to maintain a coherent plan across thousands of dependent steps for more than a day.
The paper frames the result as a test of long-horizon coding and multi-agent coordination — the same skills needed for complex software engineering and enterprise tasks. It also notes limits of the result, including the single successful first-launch run on Easy difficulty and the high number of failures along the way.
Estimated cost of the run up to the win was reported at roughly $520 at Gemini 4 Argon’s introductory API list price (higher at the standard price).
Sources
- Google DeepMind post on Xx.com/GoogleDeepMind/status/2111127991401247033
- Research paper by Maxence Faldor, Satinder Singh, and Roberta Raileanu (preprint referenced in coverage)
- AlphaSignal summaryalphasignal.ai/news/google-deepmind-s-gemini-4-argon-beats-minecraft-autonomously-in-27-hours
