Once a prompt reliably produces something you would actually send, the next step is taking yourself out of the middle of it. That is where people reach for an agent — and usually reach too early.
The difference that matters
An automation follows a fixed path: when this happens, do these steps in this order. Because the path is fixed, you always know where it broke and what it will cost.
An agent decides its own steps towards a goal. That is genuinely powerful when the work varies every time, and genuinely hard to trust when it doesn't. Every extra decision you hand over is another place the result can drift without anyone noticing.
The three-part test before you build anything
- Frequency — how many times a month does this actually run? Twice is not a workflow.
- Time saved — how long does one run take you now? Under roughly thirty minutes saved per month, you are building a hobby.
- Failure cost — what happens if it produces something wrong at 3am and nobody checks? If the answer involves a customer or a number in a report, keep a human on the last step.
Start with the boring version
The most useful first automation in almost any job is a fixed three-step chain: something arrives, an assistant processes it with a prompt you already trust, the result lands somewhere you will see it. Zapier, Make and n8n all do this well, and the differences between them matter far less than getting one running.
Here are the numbered steps I currently do by hand: [list them]. For each step, tell me whether it is deterministic (always the same) or judgement-based, and propose where the handover to a human should sit.
Try this today
Write your working prompt down as numbered steps and mark the one step that still needs your eyes. That mark is the boundary of your first automation.