State management and cascading failure loops are definitely the hardest part here. Once one agent hallucinates an output, downstream agents tend to amplify the error rapidly instead of catching it.