The First AI Agent I Actually Kept

I have built a lot of AI agents that did not survive contact with a real week.

Most of them failed the same way: they were impressive once and useless by Thursday. They needed babysitting, or they answered a question I only had a single time, or they cost more attention than they gave back.

One survived. It is boring, and that is the point.

What it actually does

Every morning it reads three things — my inbox from the last 24 hours, the orders table, and the site's error log — and writes me a short brief: what needs a reply, what sold, what broke. No dashboards. No chat. One message.

That is it. It does not decide anything. It removes the twenty minutes I used to spend deciding what to look at.

Why the others died

1. They had no owner. If nothing breaks when the agent stops, nobody notices when it stops working. My morning brief has an owner: me, at 7am, waiting for it.
2. They needed perfect input. Anything that only works when the data is clean will not run twice.
3. They tried to be smart. The useful ones summarise and route. The clever ones hallucinate and get switched off.

How I would rebuild it in an hour

Pick one thing you check manually every single day. Write down exactly what you look at, in order, and what makes you act. That list is the prompt. Give the model read access to those sources, ask for one short brief in a fixed shape, and schedule it.

If the first week's briefs are boring and correct, you have a keeper. If you find yourself opening the sources anyway, delete it — it did not remove the work, it added a step.

If you want the long version

The AI Agents Course walks the whole thing end to end: the four parts of an agent, the six patterns worth building, and the part where most agents fall over. It is $19 and it is the version I wish I had before I wasted a month.