For cross-cloud agents, I'd benchmark one complete business task, not just a query. A cheap Lakehouse read can become expensive when the agent repeats it twenty times across regions. Trace the data movement per completed task before choosing where it runs.
