Design multi-region active-active for a stateful agent platform. Where does the state model bite you?

The discriminating follow-up is 'a region dies mid-run, what happens to the 40-step agent that was halfway through calling the customer's API?' Candidates who say 'it fails over and resumes' without a checkpoint model are bluffing. The answer that scores: durable step log plus idempotent tool calls means resume-from-checkpoint in the surviving region; without that, you replay side effects and double-charge someone. The second trap is proposing active-active for write latency when the real driver is availability, then eating multi-master conflict resolution you did not need. Most teams want active-active reads with region-pinned writes, not true multi-master.