The immediate problem with the EFF's proposal is that there are lots of coding languages out there, at varying levels of abstraction and convenience. Do you have to submit a program in C, or can you submit a program in Matlab, or maybe is a UML diagram enough? In fact, it can be argued that the formal lawyer-speak that is used for describing patents today are already a form of executable code, that is intended to run on a patent examiners' hardware.
Then there's the question of what should the outputs and inputs look like.
If say I patent a method of optimization, ie to find the maximum of a function.
Should I write a small test program that uses my method to find the maximum of the curve y = x - x^2 ? That's a trivial problem that merely _illustrates_ one trivial application of the claimed method. It doesn't even _prove_ that my claimed method of optimization is bug free, since it might well give the wrong value for some other function that I conveniently didn't code for. Now think about how a patent on face recognition could be coded up...
No, patents are simply a bad idea, and should not be granted, period. Any interim suggestions to improve the signal to noise will only encourage the pro patent lobby. If requiring code turns out to improve patent quality just a tiny bit, they'll use that to claim victory: "see, patents aren't evil, the rules just have to be tuned a little bit".