I recently have pivoted to mainly working in either Claude of Codex on the command line, because their ability to run autonomously for long, extended periods wasn't possible in the turn-based Xcode intelligence features.
That, plus the fact that the agents could run command-line utilities and just generally be more "agent-y".
My way of working has really shifted this year. I've been experimenting with LLMs for coding for years now, but the models released in the past year really feel like they've hit a point where they write code much closer (sometimes indistinguishable) from how I would write it.