Evaluating Partial Message Monitoring of LLM Trading Agents: Linguistic Evasion Without Price Collusion

AI agents that set prices for competing firms could coordinate against their customers, and one defence is an overseer that reads their messages. Because reading every message is costly, real monitors may read only a sample. I tested what LLM seller agents do when they know how much of their communication is monitored, using the double-auction market of Agrawal et al. (2025) and measuring coordination separately in what sellers say and what they do. With Qwen2.5-14B agents (five runs per condition), monitoring changed what sellers said but not what they did. Under 100% monitoring, sellers almost stopped mentioning prices (from 50% to 1% of messages, p = 0.01) but kept using coordination language, in polite or compliance-themed form. The same overseer, applied after the fact, flagged 53% of rounds of unmonitored talk and none of the monitored rounds (p = 0.048): the rewording kept sellers below its detection threshold. Prices stayed at the no-talk level in every condition and at three model sizes (7B, 14B and 32B). Text monitors can therefore be evaded by simple rewording, and an exploratory check based only on prices caught coordination that the text monitor missed, so oversight of agent markets should check behaviour as well as messages.