Google DeepMind has released an experiment that makes people rethink the boundaries of "AI collaboration." 4 Gemini4Argon general-purpose coding agents, without any roles assigned to them by humans, worked out the division of labor among themselves using only the in-game chat channel and a shared note, then dove straight into a randomly generated Minecraft survival world. The hard-fought battle lasted 27 hours and 14 minutes, and they ultimately beat the game, taking down the Ender Dragon.
These "teammates" never touched a keyboard or mouse from start to finish. Their real hands and feet were lines of code they wrote themselves—calling the Mineflayer library to translate their intentions into executable in-game actions. Over the entire challenge, they produced a cumulative 2,783 programs, effectively crafting tools on the fly as they fought. Most striking was their resilience: they died a total of 605 times along the way, yet never once quit or gave up; instead, through repeated trial and error, they figured out on their own an effective tactic for destroying the End Crystals. In other words, the key strategy that defeated the boss wasn't taught to them by humans—they worked it out themselves.
The results also stand up to scrutiny for being clean. DeepMind arranged two layers of auditing—an independent judge agent and human review—and confirmed there was no cheating whatsoever throughout; meanwhile, the communication efficiency among the four agents was astonishingly high—over the entire 27-hour battle, only 229 messages were left in the chat channel. No fluff, no squabbling—just the essentials clearly stated, and back to work.
The most intriguing aspect of this experiment lies in how it diverges from past multi-agent research. Previous similar attempts typically fixed each agent's responsibilities in advance and had them cooperate according to a script; this time, the 4 agents relied entirely on their own磨合—harmonizing on their own and dividing responsibilities themselves, sustaining collaboration without a preset architecture over a period of more than a full day. DeepMind believes this capability for "architecture-free, long-horizon, multi-agent autonomous collaboration" can transfer to fields such as coding, scientific research, engineering, and even robotics. In other words, this is no longer a tool whose workflow is arranged by humans, but a team that can assemble itself and talk its way through hard problems—AI is evolving from a "you command, it executes" model toward a "you set the goal, it builds its own crew" model.