In contemporary textbooks on differential geometry, the definition of smooth manifolds is given in a (IMHO) awkwardly obfuscated way, by saying that a smooth manifold is a topological space endowed with an equivalence class of compatible atlasses. Why does it not suffice to define a differentiable manifold as a topological space together with a single, not necessarily maximal, atlas?
If I'm not mistaken, when you proceed by defining smooth maps between manifolds in the usual way, you end up with a category which is at least equivalent to the usual category.
One may argue that the definition of a smooth manifold in terms of a smooth structure has the advantage that two manifolds on the same underlying set which are diffeomorphic by means of a diffeomorphism which is simply the identity on the underlying set, are in fact the very same object, but I don't see why this should be a technical advantage.
If one makes the definition as in your first paragraph one obtains "too many" objects. On any manifold that carries at least one nontrivial differentiable atlas there'd be infinitely many different such atlasses (just add or remove a few maps), resulting in inifinitely many differentiable manifolds on the same underlying topological manifold [EDIT: For a concrete example see Branimir Çaçiç's comment]. Thus results that on certain topological manifolds there are only so-and-so many differential structures could not be formulated with the language of differentiable manifold. If taking a single atlas would make sense at all then it would be a maximal one. But that would make certain arguments more involved as well (and one would have to show the existence of a maximal atlas above an atlas one can easily describe).