A paired experiment across 86 runs tested two ways to structure a team of AI agents working together. One structure put a single coordinating agent in charge of assigning work to others. The other let agents pass work back to whoever was best placed to handle it, including back up the chain. The flat, loop-back structure outperformed the top-down one on complex tasks. It also cost less per run, with a measurable effect size. The gap came from how each structure handled failure and ambiguity. In the top-down design, a blocked agent could only wait or report upward. In the loop-back design, any agent could redirect work to a better-positioned agent. That cut idle time and avoided repeated restarts. The cost saving came from fewer wasted steps, not from doing less work. For teams building products that use several AI agents together, the finding challenges a common default. Many teams assume a single coordinating agent at the top is safer and easier to reason about. The experiment suggests that safety and transparency are design problems separate from structure. Rigid top-down control trades away performance without buying the oversight it promises.