A friend of ours, Walter Chomak, was chatting the other day about OpsMgr I/O and that he had just finished working on his performance model (not sure you ever really "finish" these things, though). The cornerstone of the model is that regardless of the number of agents, the I/O is a 85/15 split. With 85% reads, if your average disk sec/read is greater than 50-75ms, you will observe latency in the console. This will hold true for 100 agents or 6000 agents (of course, as agents increase [load], keeping those times down requires better disk subsystems.)
Walter modeled this in many sizes of environments, ranging from 500 to over 7000 agents. He has published an article on his blog, OpsMgr 2007 I/O Considerations, that you will definitely want to read. Check out http://wchomak.spaces.live.com/blog/cns!F56EFE25599555EC!610.entry.