In 2003 CERN helped write a config management system, called Quattor, in order to manage the compute grid. At the time, our scale made our problem unique, and that there wasn't any configuration management solutions available that suited us. In 2011 we looked again, and concluded that if our use case really had been all that special it wasn't anymore, and there were other widely used mature tools available, such as puppet. We have started an ""agile infrastructure"" project, which aims to use industry standard tools, including puppet, to solve what should be common problems. This talk will cover how we have traditionally done configuration management, the experiences of the pilot of our new infrastructure, and looking forward to the challenge of scaling this out to the thousands of physical machines in our computer centre and the VMs that they host.