Google DeepMind orchestrated an experiment to probe the frontiers of AI collaboration. In this endeavor, four Gemini4Argon general-purpose coding agents, devoid of any human-assigned roles, engaged in autonomous negotiations and task divisions via in-game chat and shared notes within the Minecraft survival realm. They dedicated 27 hours and 14 minutes to crafting their own code, leveraging the Mineflayer library to execute game actions. Ultimately, they triumphed over the Ender Dragon, marking the completion of the game. Throughout the experiment, the agents collectively authored 2,783 programs, persevered through 605 deaths, and independently devised potent strategies for demolishing Ender Crystals. Both agent and manual audits verified the absence of any cheating, with the agents communicating efficiently and generating a mere 229 messages. This experiment shattered the constraints of prior multi-agent research, which necessitated pre-defined roles, achieving sustained autonomous collaboration sans a pre-set framework. This proficiency can be harnessed in domains such as coding, scientific research, engineering, and robotics, heralding the evolution of AI from a mere executor of human directives to an entity capable of autonomously forming teams to accomplish tasks.
