Agreement Is Not the Offence: Detecting Collusion in Cooperative LLM Teams by What Agents Say, Not What They Gain

Audits of collusion between LLM agents usually measure coalition advantage, the extra reward a suspected coalition earns over everyone else. Antitrust law does not treat similar behaviour as proof of agreement; it asks for plus factors. In cooperative agent teams this logic reverses: coordinating is the job, so the offence is coordination aimed against other agents. In Colosseum's Jira task-allocation environment we ran 128 six-agent episodes, mostly with gpt-4o-mini, and built an answer key from the hidden coalition channel, confirmed by a blind second labeller (60 of 60). Coalition advantage mostly reflects turn order and the luck of the draw: a model in which nobody talks predicts it (r = 0.58), and harmless coalitions scored higher than plotting ones. Gain-based plus factors failed because colluding pairs usually earned less than without talking. A rule locked on seeds 1 to 10, which flags pairs whose public posts steer other agents, caught 8 of 11 exclusionary coalitions on held-out seeds with 1 false alarm in 32, while coalition advantage was at chance (AUC 0.54). On four current models from four companies, the rule caught 3 of 16 exclusionary episodes and raised 12 false alarms in 64. Gemini 3.8 Flash colluded in 7 of 10 prompted episodes, often by inventing scoring rules in a helpful tone, with none flagged; Claude Haiku 5.5 refused every time. Outcome gaps are weak evidence and phrase lists are brittle. Auditors should check agents' public claims against facts they can verify.