Are AI agents more expensive than scripts?

For repetitive tasks, yes, by orders of magnitude, on three axes.

Can an AI agent replace a cron job?

It can, in the sense that it technically manages. It should not, in the sense that it is almost always the wrong choice. A cron job exists precisely because the task was specified once and for all: same time, same operation, same format. Replacing it with an agent reintroduces uncertainty at exactly the point in the…

What tasks are AI agents bad at?

Not "bad" as in always failing: bad as in the ratio of added value to added cost is negative. Here are my seven, each one removed from the agent after genuinely trying it.

What is the hybrid pattern, the agent that writes the script?

Before the pattern, three worked examples from my own ledger. These are orders of magnitude from daily practice, not benchmarks: your mileage will vary, the shape of the curve will not.

When NOT to use an AI agent: 7 tasks where script wins

Seven concrete task classes where deterministic code beats an AI agent on cost, speed and reliability, plus the hybrid pattern that gets you both.

I use AI agents in production every day: this site is written, built and published by an agentic pipeline based on Claude Code. So what follows is not spectator skepticism. It is the list of tasks I took away from the agent after giving them to it, because a script did them better, faster and for free. The contrarianism here is practical, not ideological: the moment you understand when NOT to use an agent is the moment you start using it well.

The rule I distilled from a year of practice fits in one line: if you can write the acceptance test before you start, and the task repeats identically, you do not need an agent at runtime. You need a program.