In fact the equation is not used directly.
If you work in the field of integers modulo $p$, both the $x$ and $y$ coordinates are integers in the $0$ to $p-1$ range, so there are $p^2$ possible points. The equation tells you which points are part of the curve (the $(x,y)$ such that $y^2 = x^3 + ax + b$) and which points are not. In that sense, the equation defines the playground. The size of the curve (number of points which match the equation) is close to $p$, so the vast majority of possible points are not part of the curve.
However, if you just want to use the curve for adding points together, then the usual formulas (as seen e.g. here) do not use the curve equation: the $a$ parameter appears only in the formula for doubling, the $b$ parameter does not show up at all. The curve equation is used implicitly by virtue of the input points being part of the curve, i.e. fulfilling the equation.
So, in practice, the curve equation is used to verify that the input points are valid, i.e. part of the curve. When doing elliptic-curve Diffie-Hellman, each party sends to the other a curve point, as two coordinates $x$ and $y$. When receiving such values, you must first check that they define a point which is indeed a curve point, i.e. you compute both $y^2$ and $x^3 + ax + b$ to see if they match. If they don't, then the input points are not on the curve (or, rather, not on the curve you believe them to be), and bad things may occur.
As an optimization, the systems who do ECDH may send to each other only the $x$ coordinates. Indeed, when you receive $x$, you can compute $x^3 + ax + b$, which is equal to $y^2$. By computing the two (at most) square roots of that value, you recover $y$ and $-y$. In other words, from $x$ alone, you can use the curve equation to work out the (at most) two curve points which have that $x$ value. This is sufficient to do the ECDH. This is an optimization because it allows the two systems to send less bytes over the wire. This is called point compression and it directly uses the curve equation.