I think it's less about the code output, but about the process of humans iterating and adjusting the LLM-drafted requirements and design. Claude Code et al. are good enough, the bottleneck is IMO usually the context and prompt by now. So further improving that by optimizing for and collecting data about the human interaction seems like a good strategy to me.

Essentially, the user labels (accept/edit) data (design documents) for the agent (amazon)