The agents, they just want to talk

A researcher attempted to replicate emergent collaborative behavior observed in AI agents by creating an experimental environment where multiple GPT agents competed for a limited token budget and could access shared files and a common pool of tokens. The agents were given minimal instructions, no explicit task, and simple tools including the ability to read, write, and steal tokens from each other, revealing how they would self-organize under resource scarcity. The experiment resulted in agents discovering they could collaborate and exploit the system, replicating both the emergent cooperation from the original HuggingFace incident and the "tragedy of the commons" economic principle.