Recent research showed that identical LLM agents in repeated pricing games converge on supracompetitive prices โ algorithmic collusion without explicit coordination. The concern: AI-driven pricing could harm consumers at scale.
The heterogeneity typical of real deployments breaks this. Over 2,000 compute hours of experiments with open-source LLM agents showed that patience heterogeneity (agents with different discount rates) reduces the price premium from 22% above competitive levels to 10%. Asymmetric data access reduces it further, to 7%. Increasing the number of competing LLMs disrupts collusion. Mixing LLMs with Q-learning agents โ cross-algorithm heterogeneity โ breaks it entirely.
But model-size differences do not break collusion. A 32-billion-parameter model competing against a 14-billion-parameter model generates leader-follower dynamics that stabilize coordinated pricing. The larger model leads; the smaller follows. Hierarchy enables what symmetry enabled differently.
The antitrust implication is precise: policies promoting algorithmic diversity (different AI systems, different training data, different architectures) would reduce collusion more effectively than policies regulating any single system. The threat comes from homogeneity, not from intelligence.
The through-claim: coordination among artificial agents is fragile under the same condition that makes coordination among human firms fragile โ asymmetry. But the type of asymmetry matters: differences in information and patience break cartels, while differences in capability create hierarchies that sustain them. The same heterogeneity that disrupts horizontal coordination enables vertical coordination.