I think it's a great idea with many uses. The obvious concern is with security, and loading remote code you haven't really checked out yet. In Ruby there is a sandbox library that lets you do this kind of thing while limiting the damage a random library can do to your system. Has there already been discussion of a sandbox mechanism for Arc? As the line between client and server is blurring it might make sense to include sandboxing as a basic feature early on...
I think that some kind of signature (provided by a trusted source) should suffice. Sandboxing is propbably the most secure thing to do, but it has a few drawbacks:
1) It could potentially limit the functionalities if the loaded code (e.g. access to the filesystem)
2) Efficency, because almost every operation (such as memory accesses) should be checked to be sure the application doesn't try to get out of its sandbox.
3) Implementation: it doesn't seem to be very easy to implement :)
