| ▲ | popkkksidhd 8 hours ago | |
I do wonder about this. I’ve landed on: 1. Anything that’s boring to write - aka I’m not learning, done it before, I can quickly review - I offload to agents. And still review entirely since it’s my reputation behind it, but I’ve found it to be relatively pain free to review. 2. Anything that I don’t know what it’s gonna look like, I do my normal planning but with a lot of agent review involved up front. At the end of this, I more or less know the code that’s gonna be written and thus don’t type it myself. I still do line by line review and spend a lot of time on things I don’t understand. It still feels like the same process (wrt understanding) when I had typed and then reviewed what I typed. I still (from what I can tell) have the same mental model I did when I hand typed the code, even though for the most part I’m not. I am worried if I’m losing something, but so far it seems ok. It’s also certainly not a 10x speedup. | ||
| ▲ | yusefnapora 7 hours ago | parent | next [-] | |
You're pretty much at the top of the "vibe bobsled" [1]. Unfortunately, your lack of 10x speedup will stand in contrast to those who have descended further down the slope. Unless you have uncommonly clear-headed management, you will likely feel pressured to increase your speed, since it's so much easier to judge output volume and delivery speed than the ephemeral "quality" that you add by thorough review and understanding. Having already adopted the tooling, and especially having seen it successfully churn out the boring parts, it will become easier and easier to justify offloading the review of those parts. Then a few weeks later, some of the trickier parts become boring, and you rinse and repeat until your only remaining job function is to take the blame when the code nobody has read blows up. Not saying this is inevitable, just incredibly common. So even if you are personally able to hold the line, you should expect to be increasingly surrounded by colleagues who have gone all in. 1: https://dustycloud.org/blog/faulty-towers-vibe-sickness-and-... | ||
| ▲ | jghn 7 hours ago | parent | prev | next [-] | |
This is close to what I do. On the back half I will iterate through all the generated code other to ensure I agree with the choices, the "why" behind it all, and will often ask for or even just manually do updates and changes. The AI has done all the heavy lifting, but other than purely personal code I don't really care about everything has been vetted just as much as if I'd written it myself. | ||
| ▲ | boromisp 7 hours ago | parent | prev [-] | |
On the one hand I prefer to use software developed by people using your process. I don't trust other people's agents and prompts. On the other hand, infinite backlogs, crazy tech debt and understaffed dev teams mean a lot of the times off loading more work to agents with less human review will lead to better results on average. | ||