The link to the 50W/channel composite amp is golden. If I was going to build a class B amp I would definitely build it like this. Actually, it's sort of logical because it makes sense to drive the power stage with an op-amp IC rather than building your own driver stage. The fact that the power stage is designed to work as effectively an op-amp, is just an added bonus that makes this circuit a lot cooler.
In regards to the stability issue, yes certainly if you put two very quick devices in series and an overall feedback loop around them, it's a recipe for oscillation, since the devices will respond to the input in open-loop fashion before the feedback arrives. There are two main ways around. One way is to slow down each of the individual devices (give each one its own feedback loop, inside the overall feedback loop). Another way is to generate the feedback signal as some function of both op-amp outputs, so that there's a "quick" feedback coming from the first op-amp and a slower feedback coming from the second.
I've done a bit of this before while trying to generate stable feedback loops and integrators for sigma-delta converters. Honestly the maths does your head in a bit, but in principle the response of the circuit can be calculated by complex analysis. Given some frequency (omega), you can calculate the gain and phase shift of the amplifier in the same way you would calculate the DC gain, except using impedances instead of resistances, and obviously giving any capacitors or inductors their correct impedances, which will be imaginary and will depend on omega. You just have to make sure the gain is <1 when the phase shift is in the 90..270 degree region. The correct feedback components will ensure this.
I tend to use a trial and error approach, with either Spice or Matlab or Python with Matplotlib... so that I can easily change components and then graph the response of the circuit as a frequency-vs-magnitude and frequency-vs-phase plot. There are other plots such as the Bode plot and the pole-zero diagram, which I'm honestly not as familiar with, but can be used to read off stability information if you know how. (I try to avoid using too many different kinds of plot though).
cheers, Nick