I need more information before I can give you assistance.
How many client-nodes in your environment?
How fat are the network pipes between the sites?
Do you know of any constraints?
That will be a good start.
Primary Site around 2000 on secondary it varies. Overall all around 3500. Connectivity varies from fractional t1 to. Regarding contraints there appear to be none as far as I know.
Overall the design sounds fine - it all depends on how many remote sites there are, and how many machines per site. This controls whether or not it is worth putting a distributed repository on a remote site.
From what you're describing, I think an MA 4.6 Superagent on each site, in Lazy Caching mode, is probably a good approach.
That is the intent have full copy and updates with replication which conserve the bandwith.
OK - but if you are concerned about bandwidth usage, I strongly recommend you use lazy caching, and not replication tasks. Otherwise you could be using a lot of bandwidth unnecessarily.