I stopped asking AI to do every step.

My first instinct with agents was basically, “Keep looking. Keep checking. Keep doing stuff.” That sounds productive until you watch a system wake up, read the same information, and decide there is still nothing to do. Over and over. On somebody else's token bill.

The better move was to decide what actually starts a job. A ticket changes. A form comes in. A report finishes. A number crosses a line. Then the right little piece of the system wakes up and does its one job.

It is not an employee sitting at a desk

People can use judgment when the room is quiet. Software mostly just wastes cycles if you ask it to keep refreshing the same screen. The system works better when it is event-driven: something changes, a small task is created, and the result has somewhere useful to go.

That also makes failures easier to find. If a task ran because a ticket changed, I can trace that whole chain. If it ran because “the agent was supposed to stay busy,” good luck figuring out why it did something dumb at 2:13 in the morning.

Cheap does not mean sloppy

There is still a place for a scheduled check. Some work needs a daily summary or a periodic health check. But I want the timer to have a reason, a limit, and a record of what it found. Otherwise it is just a Roomba bouncing off the walls with a paid connection.

Events start the focused work. New or changed information creates the job.
Schedules check for things that can only be known over time. They do not pretend to be a person watching a screen.
Every run leaves a receipt. What started it, what it used, what it changed, and what it cost.

I am not trying to remove people from work. I am trying to quit spending their attention, or an AI bill, on the parts that can be handled cleanly.