The problem with this is that agents are neither going to be as reliable or as cheap to run as the software they produce that does it deterministically. I see what the author is getting at, that agents are getting more reliable, but I still think using the agent to write a tool is generally going to be a superior solution to asking the agent to do the work itself in every instance.

> agents are neither going to be as reliable or as cheap to run as the software they produce that does it deterministically

There are classes of software for which that is both unambiguously true, and will always be true. Handling HTTP requests, for example. I just find the trade-off has shifted from "you absolutely need to delegate this or program it" to also including "or just have a folder and get the agent to do it", and I'm finding this pattern pop up more and more for me.