Groups can achieve more than individuals, but multi-agent AI systems today rely on fixed rules for coordination set by designers. These static organizations often fail to adapt to the task.
The work introduces s3, a method that uses cooperative multi-agent reinforcement learning to train a set of identical language models. Each agent learns from a shared goal to decide when to work alone, when to share findings, and when to read messages. No roles or structure are given in advance.
When tested with twice as many agents as they were trained with, s3 swarms solved over 25 times more coordination problems than the best alternative. In exploration, communication cut repeated work by nearly 50%. This shows that effective collaboration can be learned, enabling swarms to scale without retraining.