AI
If the Task Is the Same Every Time, It Doesn’t Need a Model
I had a model doing the same job every morning and getting the same answer every morning. It took me a while to notice that what I’d really built was an expensive cron job.
I made this mistake myself. I had all my cron jobs running AI, calling a model every morning to do exactly what they’d done the morning before, with the same inputs and the same answer. Eventually it hit me: oh, this is the same task over and over again. It doesn’t need any AI. It needed a cron job.
So I rewrote the repeat jobs as plain scheduled scripts and kept the model only where the input really changes. Now I have 100 or so of them.
A coworker recently mentioned a startup that has an AI model running its unit tests, and my first reaction was, that’s exactly the mistake I made. Maybe they’ve got a reason I can’t see, like the model writing new tests as the code changes. But a test suite is about the most same-every-time thing in software.
Nobody really talks about this part of agent adoption. We’ve gotten so used to reaching for the model that we route solved problems through it forever. We pay for inference on every run, and we accept a little randomness in something that ought to be boringly identical every single time. That’s paying rent on a problem you already solved.
Use the model where it actually earns its keep, which is working out the task once and writing the script for it. Then take it out of the loop. A lot of the time, the best thing an agent can build for you is the thing that replaces it.
I’d bet a lot of AI workflows are really cron jobs with a bill attached.