StefanSko

That's probably all true. But I wonder if brains are just wired differently. I was never the developer that cared much about the particular shapes of any concrete IF/ELSE blocks in the code, so initially I was very bullish on using Claude Code a year ago. I realized that I'm the kind of guy who discovers the high-level architecture by writing the low-level code though. I'm not always sure what that architecture is up front, but the sheer act of writing code into a specific module sets off an internal background task probing the specific invariants of the module with respect to the architecture I have in mind, allowing me to revise and improve the design. That is all lost. All my agentically engineered projects look crude from an architectural perspective.

Very much agree. All these currently hyped workflows removing the human from the loop attribute a modularity to those agents that just does not hold up. They will always be leaky abstractions given their stochastic nature. That being said, they can be a great tool for getting past the "blank page" and just start or getting unstuck in general.