What about giving different agents varying sets of tightly scoped permissions? Is there a good way to achieve that today? If not, it might warrant the effort down the road