Instead of only hanging them evaluate the final output, you ought to also have a way to have them evaluate the process and agentic aspects in getting to said output. Claude Code outshines when you look at it end-to-end, in my experience.