Why my multi-agent workflow kept turning into transcript archaeology

I wanted AI agents to behave like a small software team. One agent implements. Another reviews. A parent agent coordinates. The system keeps going until the task is done. On paper, this sounds exactly