[saulpw]

That's what happened though? First humans built sheds, then we built 2-story buildings, then taller and taller, until we built skyscrapers. Obviously it wasn't a single structure, but we did have to evolve our thinking on how to build things, we didn't just start building a skyscraper before we built a shed.

[speed_spread]

But you didn't upgrade the shed into a skyscraper. The iterative process you describe involves a human respecifying from scratch using the knowledge developed building the previous instance and seeing it's limitations first hand. That part can't be automated, no LLM is going to challenge your design assumptions by itself. Hence people pushing agent-built projects way past what their inherent architecture should support, delivering an unmaintainable code spaghetti.

[bluGill]

You can't do that. A small bike shed is often just put some concrete blocks on the ground, and then build on top of them with wood. A correct house needs a stronger foundation at higher costs (sheds larger than bike shed are build the same way), but is still made of wood. A skyscraper is built with a very different foundation, and needs a steel frame that would not be affordable in a house. In between the two there are also building made of brick which allows building taller than wood. (and there are lots of other options with different costs - engineered wood is different)

Point is though eventually some system runs out of ability. It works different in programming from physical construction, but the concept is the same, eventually you can't make a bad early design work anymore.

[GuinansEyebrows]

to put it in another way than the other replies: you will have 100x more pushback to an arguably-necessary ground-up rewrite instead of "just add this new feature to the existing codebase", even when you (as an engineer) know full well why "just adding a feature" is probably a bad idea.