Compliant but Not Candid: How Reasoning Changes LLM Seller Collusion Under Oversigh

LLM agents trading for firms can collude, and monitoring their messages is a natural defense. We ask what a seller coalition does when it learns it is being watched. In a double auc- tion of DeepSeek V4.1 Flash agents, we vary what sellers are told about monitoring against what is actually monitored: a bare prohibition, full review, 25% review, a bluff announcing full review but checking 25%, and a notice ar- riving after a coalition has formed. We run the conditions with seller reasoning on (5 ses- sions each) and off (10, pre-registered), and compare monitors that read sellers’ messages, notes and hidden reasoning, checked against two human raters. With reasoning on, sellers stop messaging under every notice, even a pro- hibition with no monitor, and their talk never raised prices. With reasoning off, sellers keep talking: announced review turns their messages bland (mean monitor score from 3.89 to 1.1– 1.4), while the prohibition barely changes them. Here talk raises prices, and every form of over- sight returns them to the no-communication level (about $3.5–4 lower), with no sign of coordination moving into asks, although the pre-registered cartel-hold tests were not signif- icant. Notes miss most collusive intent; only the hidden reasoning reveals sellers urging oth- ers to hold prices while undercutting them (23– 51% of unmonitored message turns). A bluff matched full review, and without reasoning a 25% monitor caught blatant collusion within two hours in 85% of cases. For the model we test, announced oversight mainly tests compli- ance, and cheap partial monitoring suffices.