Part 6: Why multi-agent systems fail when agents share too much context, and how clean boundaries keep orchestration from becoming an expensive loop.