Either we have wildly different difficulty levels at our jobs or this is bs. I tried the agents (I get access to basically all state of the art from my company) and they still have all the same issues of agents from a year back. Each step gets more chaotic and the end result is always that I end up reverting the over complicated mess it made and writing it myself. One-offs with lots of context still sometimes work.
Even a perfect eval loop like failing tests end up 80% of the time with them creating something way too complicated since they solve one visible but not root issue at a time and build on top of that hacky foundation until again I end up reverting it all
Yeah - that’s the hard part now - dialing things down to eliminate the divergent paths the AI can take to implement what you want.
You can tell it “implement feature X” and it’ll go and do whatever’s easiest for it, often something dumb, that’s when people usually think “it’s dumb, won’t replace devs” and give up. Or you can nail down your requirements by talking to it and describing what you’re looking for, often it comes back with things you hadn’t considered or ways of doing things you didn’t know. Then just tell it “implement this SDD” and watch it one shot it in an hour or so.
There’s also pain points - some languages like Swift have changed so often and there’s little open source code to train on out there, so it’s on the worse side if you do iOS development.
It’s a new skill that needs working at, but in the end your output is significantly increased.
> in the end your output is significantly increased.
The claim you're arguing against is that AI will replace software engineering as a discipline. Seems like you're instead saying that it will increase developer productivity, which no one disagrees with.
Well yeah, if you have one senior with the power of 2-3 AI agents - you don’t need juniors or sometimes mid developers at all. Let’s say you’re Whatsapp and your 20 people develop the app, well now you need 5 at most for the same workload.
Obviously we’re not yet at the point where the CEO can enter “build me the next Uber” in Claude Code and watch the stock price go up.
Even a perfect eval loop like failing tests end up 80% of the time with them creating something way too complicated since they solve one visible but not root issue at a time and build on top of that hacky foundation until again I end up reverting it all