Prompt, automation, or loop?

The most valuable skill in this whole course is knowing which of three things a task actually needs: a single prompt, a fixed automation, or a looping agent. Reaching for a loop when you do not need one costs more, is less predictable, and adds risk for no benefit. Knowing when not to loop is a real skill.

Use a single prompt when the task is one-shot and you can check the answer yourself: draft this email, summarize this text, explain this concept, rewrite this paragraph. There is nothing to iterate on and no action to take — one good prompt is the whole job. Do not wrap this in an agent.

Use a fixed automation when you can write down the exact steps in advance and they do not change: "when a new file lands in this folder, rename it and move it here," "every Monday, export this report." The steps are known, so a deterministic automation is more reliable, cheaper (no AI thinking every run), and easier to audit than an agent. If you can spell out the rules, you do not need a loop.

Use a looping agent when the task is multi-step and you cannot know the steps up front, because the right next move depends on what the last step turned up: research that follows one finding to the next, a coding change that must actually compile and pass tests, triaging messy free-text requests where the right action depends on the content. This is exactly where the loop earns its cost — it adapts to what it discovers.

The decision in one line. Can you check it in one shot? Use a prompt. Can you write down every step? Use an automation. Do the steps depend on results you cannot predict? Use a loop.

The best real designs are often hybrids. Frequently the smartest build is a predictable automation with a single AI or agent step inserted only where judgment is genuinely needed — for example, an automation catches every incoming email (predictable), one AI step classifies the intent (judgment), then fixed rules route it. You get reliability where you can and looping flexibility only where you must.

Recap. Match the task to the tool: a single prompt for one-shot work you can check yourself; a fixed automation when you can write down every step in advance (more reliable, cheaper, auditable); and a looping agent only when the steps depend on results you cannot predict, so the loop must adapt as it goes. The one-line test: checkable in one shot means prompt, fully specifiable means automation, unpredictable next steps mean loop. And the best real-world designs are often hybrids — predictable automation with an AI or agent step only where judgment is truly needed.