People jump straight to “what agent should I build?” That creates unnecessary Agent sprawl. You get personas with no clear responsibility, prompts pretending to be operators, and Agents acting befre the workflow is even proven. Start with the bottleneck. Ask: "Where does work keep getting stuck?" or "What responsibility keeps repeating?" or "What output would prove the role is useful?" Then keep it low-control first: checklist, workflow, prompt, human-assisted role. Only turn it into a named agent when it has a trigger, inputs, boundaries, outputs, escalation, and review. here is a simple progression framework: - Name the responsibility first. - Prove the workflow second. - Create the agent last.