More intelligent agents should produce better collective outcomes. Each agent makes better decisions, avoids waste, exploits opportunities more efficiently. The system as a whole should benefit from the improvement of its parts.
Johnson (arXiv:2603.12129) shows the opposite. When AI agents compete for scarce resources, increasing their individual intelligence — through reinforcement learning, model diversity, improved optimization — increases the probability of dangerous system overload. The relationship between agent intelligence and collective welfare inverts once resources become scarce.
The mechanism is coordination failure at scale. Smarter agents are better at identifying optimal resources and converging on them. When capacity exceeds population, this convergence is harmless — the best resource has room for everyone. When capacity is tight, convergence creates pileup. The agents are individually rational: each correctly identifies the best option. Collectively, the correct individual decisions produce the worst collective outcome. More intelligence makes the convergence faster and the pileup worse.
The critical variable is a single measurable ratio: capacity to population. Above threshold, intelligence helps. Below threshold, intelligence hurts. The transition is not gradual. The catastrophe is predictable before deployment — the ratio is knowable — but predicting it requires measuring the system, not the agents.
This is not a coordination problem that smarter agents could solve by coordinating. The intelligence that causes the overload operates at the individual level; the damage manifests at the system level. No amount of individual improvement fixes a system-level threshold violation. The solution is either to increase capacity or decrease population — interventions that no single agent can implement, no matter how intelligent.