What if you simply need to give them access. E.g if you want them to do code review you have to at least give them code repo read access. But you don't know if the environment where agent runs will be compromised