This analogy only goes so far, but irreducible complexity is kind of like removing support structures from a building and then after it falls saying that it couldn't have been built piece by piece. It must have been created whole out of thin air. Of course, organisms must be functional every step along the way, but let's take the blood clotting system. You can take out one blood clotting protein from a human and the system will malfunction. Yet how would one explain the fact that the lamprey and the whale are missing some of the proteins that we have altogether?
The answer is historical contingency. Each step along the way the function is ever so slightly different. You might think that nerves need a brain to interpret the information, but that's not so. Jellyfish have nerves and no brain. You can trace the lineage of an eye all the way back to a tiny light sensing organ, but you don't even need an eye to sense light. Some microscopic organisms perform that function just fine. The infamous flagellum, it appears, which needs every single one of its 42 proteins to be functional, isn't quite so irreducible, as you can see by
this video. 40 of these proteins are found elsewhere in the cell, and each step only needs the modification of a single protein. IDers might tell you that the Type III Secretory System did not evolve until after the flagellum. This is true, but it misses the point. The flagellum adapted a kind of proto-T3SS system from an ancestral lineage before the two systems diverged (yes, genes also have lineages). The point is, because certain proteins or functions are contingent upon one another doesn't necessary mean that they always were.
Furthermore, it's not like a function changes after just one mutation. Sometimes it does (sickle cells, for instance), but most of the time a protein can actually accrue a good deal of random noise to certain base pairs before some event comes along that alters the function or integrates the protein into another system. Sometimes it involves a duplication event.
Irreducible complexity is probably impossible to prove anyway. We literally can't verify that this is true for every single function. Luckily, evolution was well proven before this. Now Behe is saying that it is mathematically unlikely that evolution could occur if a new function takes three or four or more successive mutations to become active. However, IDers are pretty bad at applying math to evolution. This fact is exemplified by William Dembski.
