There are two basic ways to run a multi-agent org, and picking the wrong one wastes the whole design. Sequential is a pipeline: each agent does its step and feeds the next, in order. Hierarchical is a manager: a coordinator agent assigns work to specialists, reviews what comes back, and decides what happens next.
Sequential is simpler and predictable — use it when the steps have a natural order and don't need re-visiting (research, then write, then format). Hierarchical is more flexible and more expensive — use it when the work needs judgment about what to do next, or when steps might loop (a manager who sends a weak draft back for a rewrite). Reaching for hierarchical when sequential would do just adds cost and unpredictability.
Run the seed: take a multi-step task and decide which process fits — then justify it in one sentence by pointing at whether the steps need a coordinator's judgment or just an order.