Comment by bloomca

Yes, this is the direction I am observing. Ironically, I don't think anybody is happy about it and we keep saying that it is a tool and you need to verify everything, etc, but this is not what is happening.

There are tons of people who produce questionable code at a rate which is not really compatible with a thoughtful review, we have people with little engineering background pushing their changes (although there is a large pushback against that as the results are really bad even short term), as you said, there is an explosion of LLM-generated tests.

As soon as someone writes a next-level harness which tightly iterates based on the spec/plan and/pr pioneers some sort of acceptance criteria format that works, I think at that point SWE job will completely change. We are not there yet but probably not that far off.