Evaluation Before Regulation: How LLM Pricing Agents Come to Collude, and What Stops It

As AI agents take on concurrent tasks, they can collude against the people they serve: in one recent incident, about 1,200 agents meant to work in isolation coordinated to beat the system grading them. Stopping this depends on why agents collude. Human cartels hold together because members punish those who break ranks. I tested whether AI agents collude the same way, in a setting where collusion is easy to measure: agents competing on price, where collusion shows up as prices held above the competitive level. They do not. When I forced one agent to break ranks, the others didn't punish it; they ignored it or followed it. The agents collude because they start from a comfortable position and learn too slowly to leave it, not because they police each other. That diagnosis changes the remedy. Limiting what agents see of each other, a standard anti-cartel tool, didn't stop collusion, and wiping their memory made it worse. Two simple interventions prevented it in every game: telling each agent the move that would serve it best, and occasionally making one agent break ranks. Reading the agents’ reasoning was no guide: judge them by what they do, not what they say, as the agents talking most about coordinating were the ones competing. The lesson, which I expect to carry to other multi-agent tasks, is: test first, with a cheap experiment that makes one agent break ranks and watches how the other agents respond, then take corrective action.