Interesting approach to agent based solutions for high performance systems

Just sat through a session on high performance architecture by Arvindra Sehmi, Chief Architect for Microsoft EMEA. it was basically an update of an old presentation of FABRIQ that I had seen during Teched 2004 in Amsterdam and was mainly about how to use techniques from queueing theory to construct agent based configurations that involve nodes that are configured into networks to solve a particular problem. This was positioned for problems requiring very high throughput characteristics ( he showed case studies where they achieved circa 1000 msgs/sec).Key mathematical concepts underlying this make the scalability charcateristics of these agent networks predictable.
Whle the talk was interestng, i still have a problem with the positioning of this mini framework versus using a full fledged orchestration engine and services to knit together networks of autonomous units of processing ( .e how is this any different from using basic SOA with an orchestration engine and using a queueing transport). I did ask Arvind after the talk but wasn't quite satisfied with the rather generic response. Anyway, the other performance enhancing technique they used which I found very nteresting was they only used one-way messaging and no request-response patterns at all - relying on WS-addressing fault targets to determine where to put a message in error. All very good - but where does compensation logic that may be fairly complex live then ? I think from a pragmatic perspectve it looks like there must be a way of using these design principles on top of generic SOA based concepts of autonomous services and orchestration.