I wrote up some thoughts about the challenges of verifying that AI agents write the code that you want them to write