Coding Agents Can Produce Working Code, but Software Design Still Depends on Engineers

I'm still stuck at "Can you get it right?" What am I doing wrong? It's rhetorical, but I'd take advice.

What exactly is "it," and which languages are you using?

I've had some success on small, limited-scope greenfield projects. I'm terrified to use an agent on the repository my job actually cares about: it broke everything when it was enabled by default. Ask mode is wrong about half the time, and once it goes wrong it compounds. Dave Plumber said the spec for his cross-platform task manager was 107 pages, so maybe it needs a literal novel.

Generated code usually leaves directory structure, interfaces, and state management as a haphazard mess, even with frontier models. Models make unspecified calls about error handling, often wrongly. Without a full test suite and strong type checker, I have to review every line and apply my architecture experience anyway.