Can independent AI pricing agents collude without communicating?

The chapter flags, as "a remarkable finding," that "independent AI agents are able to collude on supracompetitive prices in repeated price-setting games," citing Calvano et al. (2020) and Fish et al. (2024). It adds that the pattern is not confined to the lab: "There is also real-world evidence from the 2017 introduction of algorithmic pricing into Germany's retail gasoline market (Assad et al., 2024)." The claim is framed inside a broader observation that "price setting on digital platforms is already driven by algorithms" — so collusion among AI pricing agents is presented as a live market-structure risk already playing out, not a hypothetical for some future agentic economy.

The chapter does not itself walk through the mechanics of how the collusion arises — it cites Calvano et al. and Fish et al. for that — but it frames the result as a market-power question. Section 2.2, "Prices and market power," sets up AI agents' broader effects on prices: biased recommendations, lower search costs, or distortions to the shape of the demand curve. Section 2.4 then narrows to the case where the AI agents themselves are the price-setters, independently optimizing against each other in a repeated game, and nonetheless converging on prices above the competitive level without needing explicit coordination or communication between the firms that deployed them.

This supplies the antitrust-relevant counterpart to Why do AI-delegated firms stop exploring new business models?: both describe AI-driven market participants settling into a stable arrangement that sustains above-competitive profits, one through strategic self-confirmation at the firm level, the other through tacit price coordination at the pricing-algorithm level. It also runs the opposite direction from Can competitive pressure force unsafe agentic AI adoption?: there, competition among firms pushes behavior down toward unsafe shortcuts; here, competition among pricing agents pushes behavior up toward supracompetitive prices — a reminder that agents competing does not predict a single direction of harm, since the outcome depends on what the agents are optimizing and how their interaction is structured. It is the collusion counterpart to the sibling note Can AI agents designed by competing principals destroy bargaining surplus?, drawn from the same source and the same section.

The chapter does not run or replicate the collusion experiments or the Germany gasoline analysis itself; it cites them as established findings from Calvano et al. (2020), Fish et al. (2024), and Assad et al. (2024) without stating the magnitude of the price effect or the conditions under which collusion emerges versus fails to emerge. It also does not say whether today's LLM-based commerce agents, as opposed to the simpler reinforcement-learning pricing algorithms in the cited studies, would show the same tendency. The implication the chapter draws, rather than demonstrates, is that antitrust frameworks built around proving explicit communication or intent may not fit a world where tacit algorithmic collusion arises from independent optimization alone.