Splitting a task across agents can help when the work has distinct stages that need different skills, such as one agent that researches a topic, one that drafts a response and one that checks the draft against company policy before it goes out. Each agent can use a narrower set of tools and instructions, which is easier to test than one agent trying to do everything.
The common mistake is adding agents for their own sake. More agents mean more handoffs, more places for an error to enter unnoticed and more cost, so the added complexity is worth it only when a single well-designed agentic workflow genuinely cannot do the job.