A paper published on arXiv on October 16th revealed that researchers had four AI agents based on Google's Gemini 4 Argon model team up to challenge the sandbox game Minecraft. Without human role assignment or game logic programming, they successfully defeated the final boss, the Ender Dragon. The four agents each controlled a game character, coordinating actions through chat and shared memories, autonomously planning tasks, and using the open-source tool Mineflayer to write, run, and debug programs. The experiment started with a randomly generated, easy-difficulty survival world. The team wrote a total of 2783 programs over 27 hours and 14 minutes, completing the objective after 605 deaths. Researchers stated that this is the first known example of an AI agent autonomously completing a game end-to-end using only a standard game interface, without requiring pre-designed tools or map seed selection by humans. The study suggests that this experiment demonstrates the potential of AI in long-term task planning, programming, error recovery, and multi-agent collaboration.
