I'm skeptical that

1. Clearly define requirements

2. Clearly sketch architecture

3. Setup code tool suite

4. Let AI agent write the remaining code

Is better price-performance than going lighter on 1-3 and instead of 4, spending that time writing the code yourself with heavy input from LLM autocomplete, which is what LLMs are elite at.

The agent will definitely(?) write the code faster, but quality and understanding (tech debt) can suffer.

IOW the real takeaway is that knowing the requirements, architecture, and tooling is where the value is. LLM Agent value is dubious.

We're just in a transitional moment. It's not realistic to expect LLM capabilities to leapfrog from marginally better autocomplete to self-guided autocoder without passing through a phase where it shows tantalizing hints of being able to go solo yet lacks the ability to follow through. Over the next couple years, the reliability, versatility, and robustness of LLMs as coders will steadily increase.