The Productivity Ceiling of AI Coding Tools

The author has the opinion, that AI coding tools have made us faster, but we're still the bottleneck because we have to be present and actively engaged. He conclude the next leap in productivity requires letting AI agents work autonomously in the background while we do other things.

While I agree with the conclusion, my experience with Gemini, Claude and Devin is that there is no way around to be engaged as a human, because e.g. Gemini usually stubs a lot of thing, even though it claims that everything is implemented (which I have to check throughoutly and require it to complete very many times), Claude is often not available or stops because it runs out of token credits, and Devin makes a lot of errors or wrong assumptions. So until those systems don't become much more intelligent and reliable, the human in the loop is an unavoidable precondition to successfully use those tools. I'm not sure yet how much time I really save; my intuition is that with all the review and debugging I save between 0 and 50% of time, but I should check that systematically some day.

How is this methodology and tool (GZA) different than say using Github Copilot and hand it a bunch of issues (prompts) to work on?