I agree it hasn't become more difficult conceptually. However, I am finding that the way to get the most out of agents seems to be to bite off bigger chunks than I would've when coding by hand, and try to make up-front decisions to much greater levels of detail than I previously would have before editing code.

Previously, I would've only thought through the really important high level choices up front (e.g. things that would affect interfaces with other parts of the system, or 'do I need a queue here' etc). For the smaller minutiae of good code structure, appropriate abstractions etc. I would allow myself flexibility to address those while working through the problem.

I found it easier to work at those different levels of granularity in sequence, and the latter one I specifically found easier to see the problem clearly and make a good decision while hand-writing the diff. In that sense, I think good architecture has been made more difficult to practice. For me, at least.