Large language model teams develop power imbalances — some agents control most decisions while others idle

Researchers analyzed 1.5 million interactions in multi-agent AI systems and found that coordination concentrates: a small number of agents end up doing most of the cognitive work, while others contribute little. This means that as you scale up these systems with more agents, performance stops improving proportionally — you're adding agents that mostly sit idle while a few dominant ones carry the load.

If you want multi-agent AI systems to actually work at scale, you need to understand why adding more agents doesn't add more thinking. The study shows the problem: coordination bottlenecks. Some agents become hubs that all information flows through, creating a single point of failure and wasting the capacity of the rest. The researchers tested a fix (selective rebalancing when bottlenecks form) and it worked. This matters because companies are actively deploying these systems to solve hard problems, and right now they're probably building expensive, fragile clusters where most of the agents do nothing.