Tracker server is a reference point for all Agents. When looking for other peers for a particular folder, each Agent attempts to connect to the Tracker and asks it to:
a) let all other peers with the same folder id know where it is - it tells its local and public addresses
b) get the list of local and public IPs:ports for all peers with the same folder id.
Tracker server is automatically installed on the same computer as Management Console and agents connect to is by default. You can setup a separate Tracker server instance and configure Resilio Agents to connect to it.
In either case it's advisable to carry the tracker out of agents' LAN so that it knows agents' public address and agents can connect to each other all right.
Administrator can configure multiple Tracker servers in Profile or in agent configuration file. Multiple tracker servers can be used in 2 modes: load balancing and high availability.
Load balancing mode forces each agent to distribute their requests to all tracker servers they know of as even as it is possible. Distribution is done on a job level. For example, we have 3 jobs (A, B, C) and 2 tracker servers available. Agent will connect to tracker #1 for jobs A and C and tracker #2 for job B. The algorithm of choosing tracker is pseudo-random and gives always same results on all Agents if they have same list of trackers. Providing different lists of trackers when Agents use trackers in load balancing will result in inability to connect. This is the default mode.
High Availability mode forces each agent to connect to all trackers it can reach for each job to get maximum information of all peers. You can configure any number of tracker servers for each Agent in HA mode, although ensure that they intersect (and able to connect) to at least one common tracker server. See here how to enable HA mode.