Adding capabilities to AI-agent populations can worsen collective outcomes

When resources are scarce, will the AI agents in our phones, medical devices, cars or drones coordinate smoothly, or interfere with one another? We study populations built from large language models competing for a capacity-limited resource. We compare increasing combinations of model diversity, simple reward-driven adaptation and programmed social sensing and grouping. When capacity is scarce, the populations with more capabilities overload far more often than a capacity-matched independent reference. Some agents also earn high rewards while the population fails. No capability is always helpful or harmful: its effect depends on capacity and on the other capabilities that are present. Our findings suggest that the capacity-to-population ratio can be an important guide in the design of systems that involve AI agents competing for some limited resource.