Implementing a Conversational Agent

We have gotten a great deal of direction on implementing conversational agents from the work of Allen and Byron.
In particular, Allen points out that there conversational agents cover a range of complexities from the "press 1" phone interfaces, to their own Train scheduling system, and beyond.
It's relatively easy to make a relatively complex conversational agent.
1. You can break conversational agents into the NLU components, the behavioural components and the NLG components.
2. We've integrated all of these using GATE (from Sheffield).
3. We are currently using off the shelf NLU components including Huyck's Plink parser.
4. The NLG is really simple so Ian Kenny just hacked something together. More advanced systems are available.
5. The real work at this stage was with the behavioural components.
It quickly becomes difficult to make more sophisticated systems with Allen and his group spending a decade developing the Trains system.
All current systems, and those in the foreseeable future, will be incomplete.
If they were complete they'd pass the Turing test.
We would like to make some advance in conversational systems, so we need a domain that is interesting but not too hard.