A very good 10 MHz ovenized crystal oscillator, Hewlett Packard 10811D, ages by up to a few Hertz per year, and has a mechanical capacitor for trimming the frequency by up to 10 Hz, and an electronic frequency adjustment by 1 Hz using a varactor. I am sure the varactor does not improve jitter, but considering that the full range of adjustment is 0.1 ppm, it also should not add that much jitter, assuming everything is well designed.

So we are talking about a reasonably high stability crystal to begin with, and a very narrow adjustment range. In an atomic clock, the feedback loop uses the electronic frequency adjustment to more or less completely remove the aging. This requires a very tiny and a rather slow acting feedback.

In the atomic clock, the output of the crystal is used as a reference for a microwave sweep generator, which then scans the spectrum of atomic transitions. The absorption peak in a cesium clock is something like a kiloherz wide, but with a good signal to noise ratio and with a lot of averaging, one can measure the position of the peak to a very tiny fraction of its width. Comparing the measured and the expected positions reveals the deviation of the reference frequency from its design value, and that is what generates the tuning feedback for the crystal. I do not know off the top of my head how slow the feedback is exactly -- it is something that one could look up in the Hewlett-Packard service manual, but I am sure it is slow enough to be irrelevant for the cycle-per-cycle jitter.