I agree about the two out of four points you make, but they only cover the style part of code. In my experience, LLMs are perfectly capable of writing code which passes credo checks and follows the project’s existing style and decisions. For example, one of my recent opersource reviews was about FerricStore, a database for caching written by non-elixir engineer. No special CI, no custom credo, all generated. And it contains perfectly readable code with a good style. Of course, there is a room for improvement, but style is not the biggest problem of the project. Critical bugs and cheating is.
I don’t understand how to stop LLMs from reinventing things, cheating, making wrong decisions. These are the things which create the most problems, not the style. I am still trying to review every line that coding agent generates, but in the end I just feel like I could’ve wrote these lines myself. Maybe thats the professional deformation, but I learn about the problem more if I write the code than if I review the generated one. I can see a self-reinforcing cycle here: “I generate buggy code → I learn less about the solution → I can’t fix the generated buggy code”, and I can see it in almost every generated opensource project I review. That’s the biggest problem with coding agents I am still trying to understand and solve for myself.
is the only point I don’t agree with. It is good to have some basic level of accountability, but extra accountability just creates more problems, while bringing very little profit. People don’t make mistakes because they think “Oh, I wont be held accountable for that mistake I just made”, they make them randomly or if they lack some knowledge. Holding an engineer accountable may make them check their solution twice or maybe even write an extra test, but it will also make them more stressed (hence increasing chance of random errors), make them contribute less, offload more work on others and implement conservative solutions instead of new and innovative ones.
What does that mean? How do you do that? I know only about progressive disclosure as a part of UI/UX design.
Thats the first time I hear about it. Can you share any articles or stories about it?
).