There are a number of aspects to setting this sort of thing up. Here's what I'd be looking at:

1. Personally I recommend the tool called "puppet", especially making heavy use of the "augeas" module. cfengine is a traditional tool for this too. You may also want to use the tool "cobbler" with this. These tools between them would deal with deployment of new machines, and are especially useful for similar machine, as your dev/test/production machines are likely to be.

2. Someone recommended using packaging; that would be my personal suggestion.

3. People have been arguing SVN vs. Git. Use only one of these :).

You'll notice that the breadth of argument has occurred because there are quite a number of areas in which various pieces of sysadmin software could be used, and you haven't been clear about what areas you need covered. I'd recommend a post that clarifies what you're already using in each of these areas.

HTH,

XML. Why use it? Well, for me, the whole thing can be answered by the word "Tree" (well, "Plex" actaully, but that's less self-evident). Once data gets beyond single items ("scalars" in scripting languages), 1D arrays, and hashes, you're looking at containing your data in two major structures; multidimensional arrays, and trees (yes, I know you can do arrays of arrays, and hashes of hashes, but it's a way of representing the same thing).

Now, there are a variety of ways of selecting data out of a tree. The filesystem globbing language is a good example. LDAP's language is also good, if a bit more verbose. ACAP seems useful, but I've never gotten into it much. But to me, the queen of all the major path-selection languages is XPath. I don't like XML that much as a tree representation, but I can live with it, for the sake of having XPath.

Anyway, I hope this gives some insight into why at least one person who agrees with you in many ways is now willing to consider XML as a reasonable solution.