It goes like this: time to time gNodeB sends in DL a MAC CE with 8 beams possible to be configured. Then in DCI time to time UE gets this TCI information to switch to the new beam (one of those 8 possible beams inside the MAC CE) at a specific moment in time. All is done based on UE reporting RSRP of best 4 beams periodically
This is to my understanding, I may be wrong but this is what it is written on sharetechnote about this.
There’s also a good explanation about BEAM REFINEMENT from 5G in bullets (public sample):
http://www.5g-bullets.com/5G%20in%20Bullets%20-%20Beam_Refinement_Sample.pdf