Stop calling them "AI agents." They're scripts with a brain.
What building agents taught me: the model is 5% of the work.
The hype around AI agents conflates two very different things: the call to the model, and the system around it. The model is 5% of the work. The other 95% is the software engineering we already know — data formats, retries, monitoring, tests.
Most agents are scripts
In most cases, what you actually want is a predictable script that calls a model at one or two well-chosen steps. Not an autonomous "agent" that decides what to do on its own.
That framing changes how you scope, test and fix. It is also why I say "automation" more often than "agent" with my clients — and why I always start by setting the limits.