This resonates with my experience in codex 5.2, at least directionally. I'm pretty persnickety about code itself, so I'm not to the point where I'll just let it rip. But in the last month or two things have gone from "I'll ask on the web interface and maybe copy some code into the project", to trusting the agent and getting a reasonable starting point about half the time.

> because models like to write code WAY more than they like to delete it

Yeah, this is the big one. I haven't figured it out either. New or changing requirements are almost always implemented a flurry of if/else branches all over the place, rather than taking the time for a step back and a reimagining of a cohesive integration of old and new. I've had occasional luck asking for this explicitly, but far more frequently they'll respond with recommendations that are far more mechanical, e.g. "you could extract a function for these two lines of code that you repeat twice", not architectural, in nature. (I still find pasting a bunch of files into the chat interface and iterating on refinements conversationally to be faster and produce better results).

That said, I'm convinced now that it'll get there sooner or later. At that point, I really don't know what purpose SWEs will serve. For a while we might serve as go-betweens between the coding agent and PMs, but LLMs are already way better at translating from tech jargon to human, so I can't imagine it would be long before product starts bypassing us and talking directly to the agents, who (err, which) can respond with various design alternatives, pros and cons of each, identify all the dependencies, possible compatibility concerns, alignment with future direction, migration time, compute cost, user education and adoption tracking, etc, all in real time in fluent PM-ese. IDK what value I add to that equation.

For the last year or so I figured we'd probably hit a wall before AI got to that point, but over the last month or so, I'm convinced it's only a matter of time.