Subagents aren't the right tool for every task. If you launch an agent for a one-step job that takes seconds, you've added overhead for no benefit, so just ask Claude directly.
Tightly coupled tasks are also a bad fit. If each step depends on a decision from the previous step, and those decisions require human judgment, agents running in sequence add latency without adding value. Interactive mode keeps you in the loop at each step.
I find pipelines go wrong most often when review is skipped. If the work requires your eyes between stages, don't automate the handoffs. A pipeline that moves from research to writing to publishing without your review of the intermediate outputs can produce confidently wrong results. Agents work best when the task is well-defined and the success criteria are objective.