Ready Cohorts: Bounding GPU Opportunity and Avoiding Host Round Trips in LLM-Agent Control