Let me toss out a couple of key phrases:
Machine reading (not to be confused with "reading machine")
Evolutionary computation
Anticipatory Systems

Those three phrases lie at the heart of much of my thinking. To dive into the "literature", machine reading leads, among other places, to the KnowItAll project and to Oren Etzioni. The game being played in their earlier renditions of the Text Runner demonstration is to build concept maps while reading stuff. The concept map behaves as a weak but evolving domain model, and models like that form predictions--expectations--which are part of anticipatory systems.

I throw in evolutionary computation for reasons of eventual trial and error "genetic recombination" of concepts and relations, looking for interesting new ideas. Douglas Lenat's Eurisko serves, in my mind, as suggestive evidence that this is a worthy path to follow.

At the architectural level, there are stocks and flows; stocks exist in a growing knowledge base, and flows are the streams of new bits of information being fed in, changes being made, and all under some sort of agent-based coordination. In my general view, the best kind of agent coordination is not that which is pre-planned trough pipes, but instead is simply facilitated through devices like tuplespace (Sun's JavaSpaces-- part of the JINI open source project comes to mind). Where scheduling is required, the Apache Hadoop project makes sense. Of course, my statements reveal a Java bias, but that doesn't rule out other implementations. Done rightly, all such implementations should be able to play nice together.

I haven't revealed anything in detail, and certainly didn't say how I think it will look; I just tossed out a few terms and sprinkled in some thoughts; the rest should emerge out of serious dialogues plus trial and error implementations.