I think the ideal scenario is usually two paired programmer using a shared set of AI agents on the same working branch together. It's an ideal feedback loop of paired planning, reviewing, building, and testing.