Messy code slows agents down, just like it slows humans down. I've seen agents struggle over maintaining messy code, where every change they make breaks something else they hadn't anticipated. They can thrash for some time trying to deal with it. So I think all the same principles apply; but with different thresholds. For example, in my own code I limit cyclomatic complexity to 4 or so. But I allow agents to go as high as 10. They seem to deal with it just fine. My hypothesis is that their short term memory is larger and more accurate than ours. So I agree with you that the principles will need a different tuning for agents than for humans. But I am pretty sure that they still matter.