What percentage of your coding agent's PRs do you actually merge without touching?

Everyone posts the wins. Nobody posts the merge rate.

We've been measuring this on real repos and the number is way lower than some demos suggest. The failure isn't code quality. It's that the agent never saw the thing that made the task hard: the flaky test, the config that only exists in staging, the reason that file is weird.

Curious what other teams see:

1. Rough merge-without-edits rate?

2. What class of ticket does it reliably fail on?

3. Has anyone stopped reviewing agent diffs line by line, and what changed to let you do that?