I do not think about LLMs as little helpers that get work done for me. But rather, as little helpers that take a lighter workload and transform it into a smaller workload. Every coding agent output represents work that I have to do. I have to look over the code, understand it, test it, etc. And almost always, I have to tell the LLM what to change and fix. So before I hand a task to the LLM, the question I have to ask is, will the work that this LLM produces for me take me more or less time than accomplishing this task on my own? For most of my work, it is faster to delegate to an LLM and verify the output. Not 100x faster, or 10x faster, but certainly faster. Sometimes it’s only marginally faster, other times substantially.

I think this is a more useful paradigm when thinking about how to incorporate LLMs into your work, whatever it is that you do. When the model is finished, there is still work to be done. I don’t think we’ll ever reach a point where we can actually trust the model to the extent that we can trust a human coworker. This is why I’m no longer afraid of the “white-collar bloodbath” that some have foretold.