Maintaining projects, keeping them up to date and fixing bugs etc really isn’t difficult for them.

They tend to bloat codebases over time. It takes someone who understands the code to know how to use them correctly, and it often takes about as long to direct and correct them as it takes to just write the code, at least based on what I’ve experienced at work.

The step I’ve seen a lot of people skip when using these coding agents is learning what the code actually does, which includes learning all the libraries and functions being used rather than looking at a call and going “yeah that seems right” because it looks like it’s doing what it should. Without that, it’s impossible to maintain a codebase over a long time, LLMs or not.