It's one thing to build a software just for your own needs. It's a completely different story when your software runs on systems that you can't really control and where you have only very limited visibility/feedback when things don't go according to plan.

If the software is just for yourself: Just add some more code for handling the previously overlooked edge case and rerun the program. Else: Deal with the communication overhead and all of the misunderstandings and potentially fix the wrong problem. Also: XKCD 1172

This is absolutely spot on. The thing that surprises me the most is why so many software engineers don’t seem to understand this intuitively. Your job does not end when the pr is merged. What you just merged becomes part of a system it is also your job to ensure continuous successful function of. Even if you’ve got a large devops/sysadmin org it’s still on you to understand how everything plugs together and what context your code is running in.

Maybe the LLMs will get there. I do suspect it’s less of a capabilities problem and more of a harness problem.

I can agree and disagree here. The chance to quickly get further than before frees up more time for deeply thinking about the nitty gritty. AI is a tool, and just like a hammer it can be used to create something beautiful or ruin a work of art.