Because it has no foresight, unalloyed natural selection is in a sense an anti-perfection mechanism, hugging, as it will, the tops of the low foot-hills of Wright's landscape. A mixture of strong selection interspersed with periods of relaxation of selection and drift may be the formula for crossing the valleys to the high uplands. -- The Extended Phenotype, Richard Dawkins, p40.
For a species, exploring the fitness landscape isn't a matter of a defined-position cursor moving with a hill-climbing algorithm. The cursor can vary between wide and specific. It isn't a single thing: It's a population. It occupies the base of two peaks simultaneously, moves up both, creating competing species as it does. One is selected, the other is not, and so the cursor has jumped. Or, if you like, each member of the population has a different fitness landscape, because it has a different situation in the population as a whole. Maybe the mechanism is that the population moves to a local peak of the fitness landscape, and then grows as the selection relaxes. The population growth distorts the fitness landscape for each member, and for some members a fitness bridge is created to an alternative maximum, which they can move to. The new maximum will become a different species, possibly even in a different niche. In that way, populations may bud species in the same space+time without being replaced themselves (removing "chains" of species moving towards some as-yet-unknown global maximum fitness), and valleys can be traversed while each member of the population still has a hill-climbing algorithm without look-ahead.
It's funny. The way the population moves from peak to peak feels like quantum tunnelling. Instead of modelling a particle as a probability wave, could it instead be modelled as a population of related (virtual) particles, that are themselves selected?