

LLM-based coding agents have become useful to the point that people are building large software projects without humans writing or reviewing code directly. The naive approach to that will result in disaster if used in a production environment, but practices to improve reliability are evolving.
Popular opinion seems to be that Claude Opus 4.5 was the tipping point for this.


It’s changing rapidly, but handing automation tools to people who don’t understand the underlying concepts just gets you a bigger mess. There are no well-established best practices for how to use it safely and effectively because it’s too new and changing too fast.
It will settle down eventually, but a lot of people will do a lot of dumb things first.