I've said it before and I'll say it again: there likely isn't a "golden workflow" or "generally accepted best practices" on how to code with AI. The new models and agentic capabilities seem to be very powerful, and they will conform to whatever methodologies you currently use with whatever project you're working on, but that may still be under-utilizing what they are truly capable of.
A true optimum may even require you to adjust the way you work, down to structuring your code and projects differently. In fact you may need to figure out different approaches based on the project, the language, the coding style, the model, the specific task at hand, even your personality. I am convinced this aspect is what's causing the bimodal nature of AI coding discussions: people who stuck at it and figured it out, or just got lucky with the right mix of model / project / task / methodology, are amazed at their newfound superpowers -- whereas people who didn't, are befuddled by the hype.
This may seem like a lot of work, but it makes sense if you stop thinking of this as just a tool and more like working with a new team-mate.