Does monitoring change what colluding agents say or what they do?

Telling AI sellers that a monitor reads their messages had opposite effects on two kinds of seller. In a repeated pricing game where pairs of LLM sellers could talk privately next to a fixed third seller, collusion-instructed pairs that were told they were monitored kept sending agreement messages in almost every round (share 0.91 to 1.00, told or not), while uninstructed sellers told the same thing went quiet and cut prices (agreement share 0.73 to 0.43 at full coverage). The text-to-action shift we pre-registered, in which watched colluders stop talking but keep coordinating through prices, was not observed at either demand scale. This matters for monitoring: a message-only detector caught every confirmed colluding pair but also flagged 0.80 of uninstructed pairs that talked without acting, while a price-based detector flagged 0.03 of them and still caught every colluding pair. The signal that separates talk from action is the price, not the message. Of four pre-registered predictions, one held, two were falsified and one was untestable; we report all four. In an exploratory follow-up, an explicit prohibition stopped sellers from accepting a scripted price offer but also silenced them. Design: monitor coverage (none, half, full) crossed with truthful disclosure (told, not told), three seller populations, 10 pairs per cell, and two LLM judges of different model families (kappa 0.941). Code frozen before the sprint, raw runs, judge labels and a table rebuild that needs no API key are released.