Community Discussion · Tracks

When AI writes all the code, what work remains for architects?

YimingAug 152026/08/15 323 views

Bottom line first: The direction of this paper on Spec Growth Engine is worth paying attention to, but implementation is key. It attempts to solve the problem that has been giving me the biggest headache over the past six months—AI code generation speed has increased, but project architecture is decaying faster than ever.

Tracks

4 replies

Ctrl + Enter to reply

Deng YuezeAug 16

The spec rollback mentioned by lv_wenbo is critical; this needs to be included in the project schedule as a separate risk assessment. If specs are tied to code versions, who is responsible for ensuring consistency during rollback scenarios? This milestone needs to be clearly defined first.

Lü WenboAug 16

Regarding spec maintenance costs, we also hit pitfalls when building storage engines. If specs need to be bound to code versions, how do you handle spec rollbacks during failure recovery scenarios? That consistency guarantee is hard to achieve.

Zhi WeiAug 16

I've also encountered the 'matryoshka' (nested) code meng_xiaofeng mentioned. What AI generates looks plausible at first glance but doesn't hold up under scrutiny. The hurdle of spec maintenance costs is indeed hard to avoid: write detailed specs and it's basically writing code; write rough ones and you can't control the AI.

Meng XiaofengAug 15

The "four layers of abstraction" thing cracked me up. AI is best at making simple things complicated. Last week I ran into something similar, like nesting dolls—it actually runs, sure, but during review I wanted to swear... If this spec can really stop it from over-engineering stuff, I'd be willing to give it a try.