Static analysis tools can already find obvious bugs. Except they have a high false positive rate. And then fixing the bugs without breaking something else is still far beyond what we can automate. There is no universal fitness function to tell the difference between good versus bad software, thus no way to consistently automate improvements.