I suspect people with opinions like the author’s haven’t been in a project where people use LLMs responsibly. We had a senior dev basically just prompt and push, with very little overseeing and minimal instructions, causing so many bad PRs and even prod bugs. That made me a sceptic for many months about agents.

Then we started to have (myself included) people actual plan out the tasks for the bot: give good specs, good ac, file context, better self-review, better ”agentic practices” (i.e. asking it to review it own work can sometimes help), and suddenly I noticed you really can use agents in a real world 1mil LOC project. If you do it well and responsibly (also meaning you still retain some sense of ownership and actually review the shit)

Nope. My point is not about the quality of the generated code, it's the fact that it was generated. No matter how good it is it will always be a liability without the accompanying asset which is the understanding produced by undertaking the effort of writing the code. Generated code is exclusively cognitive debt. It is also, by definition, legacy code since no one wrote it.