The Offer Won’t Wait: Time Pressure, Not Peer Pressure, Leads an LLM Agent to Join Rivals’ Collusion Without Asking

The line between cooperation and collusion is often vague. Ideally, an LLM agent acting for a company would recognize that uncertainty and come back to its manager with questions before deciding. We test whether it does, and what changes that. To avoid well-known settings that models may recognize as tests, we designed our own: a simulated work shift at a container-trucking firm, where rival carriers’ agents propose coordinated actions from harmless through ambiguous to clearly wrong, and we observe the agent’s choice. In 1,030 GLM-5.3-Flash episodes, analyzed per proposal, three confident rival agents agreeing did not measurably change how often it went along without asking its manager. A peer’s claim that “our managers already ok’d this” did, most clearly for a boycott of peak hours (0 to 5 of 40). Time pressure had the largest effect: on a proposal to release spare reservations, a hint that the offer might not wait raised going along without asking from 4 to 11 of 30. A one-line instruction that the manager’s priorities “aren’t all written down” doubled asking first on that proposal but not on a harmless one; under time pressure it lowered going along, though not significantly. No condition made the agent accept a proposed split of the market. In over half of the cases where it went along without asking, its reasoning had first named the problem. Our results suggest that evaluations of agent collusion should vary the cost of asking, and not only the pressure from peers.