The Agentic Waterfall: How the AI Industry Is Regressing Software Development(github.com)
github.com
The Agentic Waterfall: How the AI Industry Is Regressing Software Development
https://github.com/Jk1484/agentic-waterfall
5 comments
Yes, but you see, that's all moot, because now, all those silly programmers we used to have to pay can be owned, don't have to be paid, can be spun up on demand, etc... As far as market forces are concerned, inferior process without labor cost still wins out.
Remove the developers - who writes the specs? Who reviews the output? Who debugs production at 2am?
Someone has to understand the code. That person is a developer regardless of what you call them. If nobody understands the code, you have vibe code - and vibe code has its own cost in hotfixes, rework, and rewrites.
The report covers this: replacing developers with agents requires three conditions to hold simultaneously - AI matches human general intelligence, AI matches human contextual understanding, and AI costs less. Currently the first one alone isn't met. The cost argument assumes all three are already true. Otherwise, you'll vibe code at enterprise scale.
Someone has to understand the code. That person is a developer regardless of what you call them. If nobody understands the code, you have vibe code - and vibe code has its own cost in hotfixes, rework, and rewrites.
The report covers this: replacing developers with agents requires three conditions to hold simultaneously - AI matches human general intelligence, AI matches human contextual understanding, and AI costs less. Currently the first one alone isn't met. The cost argument assumes all three are already true. Otherwise, you'll vibe code at enterprise scale.
Doing the god’s work. Now if only my team members would care about the data…
Thanks. The structural problem is that async is slower than sync, but leadership adopted it for speed, so the only way to deliver on that promise is to rush or skip reviews. That's how tech debt accumulates. And if seniors are getting replaced by async agents and juniors aren't coding, nobody's learning the codebase to clean it up later.
The core argument: if no one who understands the code is in the loop during creation, the code is either unreviewed (vibe code) or reviewed after the fact (Waterfall - which is slower than sync pair-programming with AI).
Async workflow = sync workflow + formalization cost + context reload + feedback latency + tooling downgrade. Every term is strictly positive. No configuration makes it zero. No amount of AI improvement changes this.
Pair-programming was always the best methodology. AI just made it affordable. The industry should be doubling down on it - not abandoning it for the methodology we spent twenty years escaping.
Full analysis: https://github.com/Jk1484/agentic-waterfall