Why do I need the Evaluate[] in the code below?
cdf = CDF[NormalDistribution[0, 1], x]
Plot[D[cdf, x], {x, -3, 3}] (* Doesn't work *)
Plot[Evaluate[D[cdf, x]], {x, -3, 3}] (* Works *)
Plot can be thought of as an extended Table, with the difference that Plot gives a graphical output.
Testing your example with Table gives some insight from the error messages and in the output:
Mathematica users must get used to this. Plot gives a numerical value to x before it evaluates the derivative, resulting in something like (d f(1)) / (d1) from (d f(x)) / (dx). Evaluate forces the symbolic calculation of the derivatives before the numerical values come in. Also compare
Plot[RandomReal[], {x, 0, 1}]
Plot[Evaluate[RandomReal[]], {x, 0, 1}]
One more explanation:
Plot has the attribute HoldAll - for good reasons! Just mimic of not having this attirbute (and try):
x = 3;
Plot[Evaluate@Sin[x], {x, 0, 2 Pi}]
I suppose one might take the question to mean, Why was it decided to have Plot[] hold its arguments and not evaluate the function(s) until the variable has been given a value? It is a choice. If Plot[] (and Table[], etc.) let their arguments be evaluated, then we would have to learn to program our plotting as follows, unless we could be sure x had not already been given a value:
Block[{x},
Plot[D[Sin[x], x], {x, 0, 10}]
]
Programmers could Block[] the variables themselves instead of having Plot[], Table[], etc. do it automatically. We would get used to it, and maybe someone would be asking, Why do I have to use Block[]?
Instead another choice was made, and occasionally we have to use Evaluate[]. As Gianluca said, Mathematica programmers have to get used to this. The number of times I have to use Evaluate[] is fairly small compared to how often I would have to use Block[], so the choice seems convenient to me. OTOH, I was used to having to localize variables in C, so having to use Block[] all the time would probably have seemed normal.