[jemmyw]

The best results I'm seeing with my own work is mixing agent with my reading of the code and advising. I find it super hard to imagine losing the human view right now. Agents produce good code but they too often don't see bigger structures. They duplicate code and logic and then can't find all the copies if it needs updating, they still need a person to say "make a helper" or whatever.

[jmpz]

True, but I can't help but notice how far agents' capabilities have grown in just the past six months, so I assume they'll continue to improve to the point where I tend to, unfortunately, agree with Schmidt. I mourn it also — but I suspect 'code' will take a more representational form. So we can describe it to agents, at least in codebases that mix human and AI written code.