A good/complex feature or component typically consists of 2,000-5,000 lines of code. Currently, foundational and reasoning models can generate approximately 250-300 lines of error-free code, requiring 2-4 iterations. As the code length increases, both error rates and the number of required iterations increase proportionally.
We're still far from a future where developers serve purely supervisory roles. The lack of deep technical stack knowledge makes iterations nearly endless. The experience and iteration speed vary significantly between developers with tooling knowledge and those without it.
My theory is similar to how teams with in-depth knowledge of musical notes and tunes produce great music – the same principle applies to products, agents, and code in general. Engineering teams using AI will consistently outperform non-engineering teams in AI implementation.
Therefore, the best approach is to enhance your skills and learn foundational knowledge. This knowledge will remain valuable and make your collaboration with AI more productive, rather than trying to brute-force your way through challenges.