This was GPT 5.5 and codex. The specific model and harness isn't that important here. AI could do it. But the issue seems to be that there are some tasks where AI kind of falls over and provides poor results. It was easier, better, and faster for me to just do it myself. I have found a lot of cases where AI is great. If you have a UML diagram already, or translating code from language x to y, fixing unit test failures, generating boiler plate. But I can definitely see if people are using large amounts of AI for writing code, analyzing code, etc. that they are not actually seeing returns.