Verify, Don’t Decode: When Trusted Agents Disrupt LLM Collusion in Software Engineering and Double-Auction Markets

Colluding LLM agents can hide their coordination, so monitors that read their messages can be fragile. We tested whether trusted monitors placed inside a multi-agent system can disrupt collusion without first decoding it (“verify, don’t decode”), in two domain settings with a shared, tamper-evident evidence log. In a double-auction market, a trusted competing seller admitted once overpricing was verified raised buyers’ surplus in 4 of 5 collusive markets. Quarantining suspected sellers did not help. Publishing the competitive price dissolved an open cartel but not a hidden, steganographic one. In a pre-registered software-engineering study on a real FastAPI change, harm was too rare to disrupt, and a trusted “probe” teammate drew no response. Once teams were competent (9/10 honest teams correct), a trusted verifier tripled correct releases but also vouched for a colluding partial fix its tests did not cover. In a chain of monitors (a weak trusted inner desk plus an isolated, capable but untrusted reviewer checked with honeypot audits), an honest reviewer raised buyers’ surplus in every collusive market. A scheming reviewer facing audits reasoned about which case was the audit and withheld verdicts. Lesson: layered monitors must fail closed and match authority to reliability. The core goal is can multi-layer chain of monitors in double vacuum environments disrupt impact of colluding agents.