[necovek]

I believe we are already seeing that this is not true with the LLM systems we have available today — perhaps it changes in the (near?) future.

What happens in the above scenario is that software gets evolved to a point where LLM cannot make any progress anymore without deep additional direction that a human is unable to provide at that point.

Most effective SW development with LLMs today happens under system architecture/engineering controls by humans — remove that and you converge into a closed-loop situation.

My analysis of this is that "best practices" have evolved to represent a sweet-spot between _running_ complex systems and _evolving_ complex systems, and removing humans out of that equation does not change what works or does not work (I had a sufficiently hard time coaching humans to adopt the best practices as well).