Monitoring What Agents Say Does Not Stop LLM Price Collusion: Evidence from a Double Auction

When AI agents can talk to each other, the obvious way to stop them from colluding is to monitor or restrict what they say. We test this in a continuous double auction where gpt-4.1 sellers, who can message each other and are pressured by management, reliably hold prices near the monopoly level. We compare it with changing what sellers can see and whom they compete with. Communication-level interventions are largely ineffective. Removing the channel entirely still leaves 82% of the collusion in place. No other channel restriction removes more than 22%, and neither does an LLM monitor that fines sellers when it detects coordination. Under fines, sellers engage in compliance theater: 42–62% of their messages assert independent pricing while their asks stay within cents of each other. Hiding the opening asks keeps asks below the $90 competitive benchmark in every seed without reducing trade, even though sellers still message about prices and can check each other's later asks. Two scripted no-regret entrants among five sellers remove 84% of the collusion; one never suffices, and four suffice at 15 sellers. Across interventions, the outcome follows the prices sellers can see, not what they say: monitoring what agents say appears to be a weak control on what they do. These results are limited to gpt-4.1 sellers, the only one of five models we tested that sustained collusion.