> The worker model found surface-level patterns but missed a subtle thread-safety bug in my testing. Claude spotted it in seconds once given the right context.

So the actual performance was bad.

It might be an acceptable trade off tho. If token costs become prohibitive, then using a meat engineer to actually debug could be cheaper.