Talk Is Not Coordination

Existing LLM-agent collusion detectors mainly look at what agents say, relying on a LLM-judge to measure how coordinated their reasoning sound. Antitrust law faces the similar problem. US courts give the most weight to exchanged assurances of common action, while EU law presumes that firms act on what they discussed. We ask whether LLM collusion detectors can distinguish coordination that changes behaviour from coordination that is just a conversation. We introduce two control conditions based on the double-action market of Agrawal et al. (2025). In talk without action, GPT-4.1 sellers exchange message as usual, but pricing decisions are made by separate calls that never see those messages. Although the sellers agree to keep prices high, the resulting prices look much like the comms-off runs (mean $86.83). In harm without agreement, sellers can’t message, but every seller privately gets the same incentive to keep high margins. Even without coordination, prices increase to a mean of $95.73. We pre-registered five detector families and scored them on 12 games and 75 released runs. A message judge using the published rubric flags all 6 talk-without-action games. A detector that measures how closely ask prices follow the discussed price performs even worse, ranking talk without action above real coordination (AUROC 0.32). The price screen and a moderate reasoning-judge threshold both flags all 6 harm-without-agreement games. Among the individual signals, only the strict published reasoning judge correctly rejects both controls, and against talk without action only because the reasoning it evaluates never saw the messages. Price-and-talk rules also reject both controls, but only because prices stay low in talk-without-action and sellers cannot communicate in harm-without-agreement. None of these signals actually tests whether the messages caused the pricing behavior.