Having used a less formalised version of the process (manually managing the agent files but having the user / agent conversation structure the same), I've been getting some really good results out of the agent on some long running complex tasks. I'm seeing so many people talking about leaving agents running and completing projects end to end without any intervention, but my experience so far is that decent software still needs architecture guidance and a human sense of 'taste'.

Exactly this. "Architecture guidance and human taste" is the part that doesn't get automated. In my experience the AI is very good at implementation once the constraints are clear — but the constraints have to come from somewhere. That somewhere is still a human who knows what good looks like.