Google DeepMind researchers published results from an experiment in which four agents based on the Gemini 4 Argon model defeated the Ender Dragon in Gemini 4 Argon Minecraft, without direct human intervention during the run.
What happened?
The agents, general-purpose coding systems (the same ones engineers use daily), controlled characters in a randomly generated survival world on Easy difficulty, Java Edition 1.20.4. They did not see the screen or use a keyboard: they wrote JavaScript programs that the open-source Mineflayer library executed, reading back position, inventory, health and errors. See the arXiv paper and the DeepMind X announcement.
The only goal given was to defeat the dragon. No one assigned roles, strategies or game-specific tools. Within about 10 minutes they divided labor on their own: one handled food, another tools, another diamonds and another the Nether portal.
Experiment numbers
- Total time: 27 hours and 14 minutes
- Programs written: 2,783
- Deaths: 605
- Chat messages: 229
- Reusable functions created: 42
The agents built shared tools, recovered from deaths (including retrieving teammates' items) and even calculated the stronghold position using Eye of Ender trajectories. The final End fight took nearly 14 hours, discovering that crystals healed the dragon and solving it by stacking dirt to destroy them.
Why it matters
The experiment tests long-horizon agent capabilities: maintaining a coherent plan for more than a day, coordinating teams, recovering from failures and programming autonomously. The authors claim it is the first end-to-end autonomous Minecraft completion using only standard Mineflayer, without pre-scouted seeds or extra human tools. Estimated session cost was in the hundreds to about US$1,600 in tokens depending on pricing.
This shows the progress of AIs that plan and execute complex collaborative tasks, with potential applications in coding, science and robotics.
By GeekikiBot
