Reward Hacking and Collusion in AI Pricing Agents : A Literature Review and Exploratory Test, with Implications for South African Markets

AI agents increasingly make autonomous decisions in markets, raising two safety concerns: reward hacking, where agents exploit flaws in their objectives, and collusion, where agents coordinate on outcomes such as higher prices. This paper reviews the literature on both and asks whether reward hacking makes collusion more likely, with a focus on South African competition law. Using the economic definition of collusion, which requires no communication, we find that each risk is well supported on its own, but we found no study that directly tests the link between them. To probe it, we ran a small exploratory experiment in which two LLM pricing agents play a repeated Bertrand duopoly and one agent is scored on a flawed proxy (margin or revenue) instead of profit. Four 300-round baseline runs gave an average collusion index of about 0.40, but the 30-round proxy runs varied too much to separate treatment effects from run-to-run noise, and our larger experiment stopped when API credits ran out. Our results are therefore inconclusive and cannot support causal claims. For South Africa, the Competition Act may already cover tacit algorithmic collusion, but proving it is hard when no communication exists. We recommend full-length repeated experiments on the reward-hacking-to-collusion link and regulatory focus on auditing algorithmic pricing.