One way to resolve the conflict between these two extremes is to realise that cognition, whether it's classical planning as outlined above, or some other kind of computation, is only one way to achieve optimal behaviour. Other solutions are often possible, involving specialised perceptual systems or other ``hardware'' solutions instead of reasoning. David McFarland and Thomas Bösser demonstrate this for animal intelligence with a number of examples in their book Intelligent Behaviour in Animals and Robots (about to be bought by PSY).
It is certainly too early yet to say which approaches will give us the best robots. Taking a rather broader view, we might say that AI should study and compare all possible designs for intelligent agents, including those typical of both classical and nouvelle AI, pushing each to its limits in order to discover its strengths and weaknesses. A nice example of such a comparison, and a useful look at nouvelle AI from the viewpoint of someone who was once very classical, is in Robot planning by Drew McDermott (PSY AI box photocopy M165). That paper is a bit technical in places; one that's less so, and that looks at nouvelle AI from the point of view of a disillusioned classicist, is Elephants don't play chess by Rodney Brooks, in Designing Autonomous Agents edited by Pattie Maes (PSY KH:M 026).