We have described an implemented spoken-language dialogue system, which combines deep and shallow language-processing engines and an agenda-driven dialogue manager. We have also described a preliminary evaluation of the system. Although the limited set of data prevents us from drawing any firm conclusions, we feel encouraged in further exploring parallel shallow and deep language processing in the context of spoken-dialogue systems. There are principled cases that the shallow processor has problems dealing with, and in our limited experiment the configuration that combined deep and shallow processing was at a slight advantage relative to the one that only used shallow processing. On the other hand, our hypothesis that deep processing was more advantageous in situations where the user takes the initiative did not receive support by this experiment. Actually, the RP appeared slightly more successful both in general and on user initiatives. However, in our experiment this difference could be largely explained by the extent to which the CLE chose the wrong (grammatical) fragment from the N-best list. It thus seems that we had underestimated the degree to which output from the speech recognizer would require fragment analysis whose results might require careful selection. To be able to deal with this, we would have to improve fragmentary-analysis selection in the CLE so that decisions can be made statistically from the results of supervised training over already parsed corpora. Some work has been done to integrate this technique into our general tool for customizing the disambiguation component of a language processor [Carter1997].