At the labs, ~100% of code is now written by agents, this metric is not interesting anymore, but they are still reporting only that number. I haven't seen the more relevant one nowadays: what fraction of the code they actually read, how closely, and how their time splits between brainstorming with the agent or alone, reading code, writing specs, reviewing tests, etc.
I’d like (a) personal data points here and (b) to push for big firms to track and publish this instead of “% written by AI”.
e.g. I've always hated both writing and reading code, and since Opus 5.5 dropped I've stopped reading it almost entirely. The most crucial chunks I still read, since unreported stupidities still show up. I spend no time designing the tests or any kind of deliberate specs, relying on agents' holistic understanding of the picture after our back-and-forth on what needs to be done, not sure if it's a mistake.
My experience since Opus 5.5 is very similar to yours. Mistakes are so vanishingly rare by now that I mostly rely on its outputs without looking at the code, let alone write it myself.