I can see some merit in this idea. On a similar note, my company has GH Copilot code reviews, so I regularly generate code with Copilot using the Claude Sonnet model locally and then Copilot reviews my PRs (in addition to humans). A lot of times, the code review feedback is on point, and then I often copy and paste it back to the Copilot agent I’m running locally to address the code review feedback.

Having 2 passes of AI does improve the result, though it would quickly go off the rails without senior engineers reviewing and steering the output, not to mention putting the initial architecture in place. From my experience, I can’t imagine building anything with AI that has any reasonable amount of complexity that won’t eventually collapse in on itself without the guidance of senior engineers. Multiple AI agents working as an ensemble won’t eliminate the need for that guidance, IMO.