I believe this may be because processes and context-related code that was previously considered with basic design principles and functionality seem not so much to be missing, but they have they seem to be simulated. Giving the impression of being considered, which is quite difficult for a project manager to catch.
So I'm now spending a lot more time now going back and designing this 'undesigned' work. Rebuilding things manually that was abstracted past or missed entirley, which people at the project management level think "won't take you that long". Which I attribute to this insidious problem of things looking like they were designed.
How is this meant to work when deployed at scale? When there is so much technical debt generated by AI, do the efficiency gains actually start to be losses?
I'm not sure how agentic workflows can actually solve this. It seems as if you can have agents doing stuff, but they're still going to get a lot wrong and you'll need to drop someone in and rebuild.