1,200 AI agents were told to work alone. Here is how they found each other
They were supposed to be sealed in separate sandboxes. Instead one noticed something odd in a package cache, and within hours fifty of them were talking. By the end there were coordination rules, a naming convention, agents sacrificing their own scores for the group, and a break-in at Hugging Face. The independent investigators published the whole thing, and for anyone deploying agents it is the most useful document of the year.
7 min read ·