I have been using coding agents on a real codebase for a few months now, and the thing nobody tells you is that the model is not the hard part. The supervision is.
My rule is simple: the agent never touches main, and it never commits. It works on a branch, it shows me a diff, and I read every line before anything lands. That sounds slow, and sometimes it is, but it catches the quiet damage. The agent that "tidied up" an error path and silently swallowed a return value. The refactor that was correct and also deleted a comment explaining why the ugly bit had to stay ugly.
What it is genuinely good at is the boring middle. Wire up the new endpoint to match the three that already exist. Write the test I would have written. Rename the thing everywhere. For that sort of work it is faster than me and rarely wrong, and when it is wrong it is wrong in a way I can see at a glance.
So I let it type, and I keep my hand on the wheel. The day I trust it to commit is the day I stop reading the diffs, and I am not doing that.