Good question! The variable being proxied is inherently unmeasurable. The experimenter can either leave it out of the model or choose a proxy for it. An example of such a variable is education. A suitable proxy for it could be years of schooling.

A variable being instrumented is very much measurable, but it is endogenous. Here, the experimenter can either leave it *in* the model (and make the coefficients biased), or replace it with one or more IVs.

To make matters interesting, one could instrument a proxy variable (which is incidentally what the example in the article does - education is proxied using years of schooling, which is in turn instrumented using mother and father's years of schooling).

The key thing to remember is, proxied variables are unobservable, instrumented variables are observable.