I also struggle to reconcile these things, I guess the only way would be to try and see if it works for me, without caring too much if it works for other people on the internet. I would just like to add that, DwarfStar's author (Salvatore Sanfilippo) is a strong supporter of the idea that you shouldn't read the code, and he says he never read DwarfStar's code. And still, it seems that this project is much more than PoC and actually both usable and useful for people doing local inference (I didn't try it myself, but I saw a lot of positive comments about it). Could be that the crucial point is in how we use those models: instead of giving it a general goal (e.g. build me an inference engine) Sanfilippo, being an experienced programmer, kept pointing the models in the right direction. He also read the papers related to the models he was programming support for in DwarfStar, so that, when he worked on optimizations, he knew what should be done instead of prompting a general "please optimize this". So, I would say that, if you let the agents work on a "feature by feature" basis instead of trying to on3-shot things, you get much better results. Could also be that, by attempting to one-shot large projects, the model starts coding badly due to context window exhaustion.

Sorry for the not so well written comment, I was just throwing in some ideas.