What strikes me reading this thread is how many of you have independently landed on the same conclusions: human accountability for generated code is non-negotiable, architecture has to drive generation (not the other way around), traceability matters, and you can't maintain what you don't understand.
These aren't just gut feelings — they're actually the core of something a few of us have been trying to formalize: the Agile Vibe Coding Manifesto (https://agilevibecoding.org). It explicitly builds on top of the original Agile Manifesto to address exactly this environment.
The four values it proposes — accountability over anonymous generation, traceable intent over opaque implementation, discoverable domain structure over scattered code, and human-readable documentation over implicit knowledge — map directly to what experienced engineers in this thread are describing: staying deeply in the loop, reading the core code, and defining architectural boundaries before letting AI generate within them.
It doesn't reject AI-assisted coding. It tries to answer: what disciplines do Agile teams need so that AI-generated code doesn't become a liability? Worth a look if this resonates.