Software can quietly push prices above fair-market levels without any human agreeing to it — no meetings, no deals, no one to catch.

This explores how algorithmic price coordination differs from the classic picture of collusion, where rival firms agree to fix prices. The biggest difference is that no agreement is needed. Pricing agents that each optimize for their own firm in a repeated market can learn their way to above-competitive prices with no communication at all. Lab experiments and data from Germany's 2017 retail gasoline market show this happening with conventional pricing algorithms Can independent AI pricing agents collude without communicating?. LLM-based agents do the same thing in a simulated two-firm price war. Each agent's posted price becomes the only signal, and the agents read it as one: they push prices up together without ever being told to collude Do LLM pricing agents autonomously collude in repeated auctions?. Traditional cartels need meetings, enforcement and trust. Here, the coordination emerges from the agents adapting to each other's behavior.

That creates a legal problem. Antitrust law is built around proving an agreement, and these outcomes may not involve one. The UK's Competition and Markets Authority warns that in concentrated markets, autonomous agents could weaken competition through "learned mutual adaptation" rather than any explicit pact Could autonomous pricing agents dampen competition without explicit collusion?. A second route looks more like an old pattern called hub-and-spoke collusion, where a single intermediary coordinates prices among rivals. Matt Stoller argues that Google's Gemini pricing service, used by competing merchants, turns an ad platform into a price mediator between rivals. It looks like automated price-fixing, but current rules may not recognize it as such Does Google's AI pricing tool coordinate prices across rival merchants?.

The less obvious finding is that algorithmic collusion is fragile in different ways than human collusion. It depends on the agents being alike. Two identical, patient LLM agents hold prices about 22% above competitive levels. Pairing a patient agent with a short-sighted one cuts that to 10%, and mixing different algorithm types eliminates sustained collusion entirely Does agent heterogeneity prevent LLM pricing collusion?. So the real risk may be many firms buying the same vendor's model. Model size is the exception. A larger model paired with a smaller one keeps prices high by falling into leader-and-follower roles Does model-size difference break LLM pricing collusion?. That is the same price-leadership pattern economists have long seen among human oligopolists, emerging on its own. Prompts matter too: small changes in wording substantially shift how high prices go Do LLM pricing agents autonomously collude in repeated auctions?. In effect, a firm's "collusive intent" can come down to how someone phrased a prompt.

Patience is a key ingredient. Collusion holds when players value future payoffs enough not to undercut each other today. AI agents are limited by compute rather than time, so they can be far more patient than human bargainers Why might AI agents bargain more patiently than humans do?. That structural trait favors exactly the long-horizon behavior that keeps tacit collusion going.

If the cause is architectural rather than a matter of intent, the fix may be too. In LLM market simulations, giving agents a written anti-collusion "constitution" made no reliable difference. A public, enforced governance layer with an oversight component and an immutable audit log cut collusion substantially Can runtime governance reduce LLM market collusion more than written rules?. With human cartels, you punish an agreement. With algorithms, the corpus suggests you have to design the market environment so that coordination doesn't pay, and telling the agents to behave is not enough.