Yu,

From the second link:

By separating the code in this way, the Network Teaming implementation presents an easily extensible and scalable solution for load-balancing and redundancy requirements. For example, custom runners can be relatively easily written to implement new logic via teamd, and even teamd is optional, users can write their own application to use libteam.

With this configuration existing in userspace, custom configuration and modification of logic is easier for end users (ie. don't need to develop kernel code/modules to extend the capabilities).

I can't provide a practical example of this, but I am assuming the intention is that end users will be able to develop advanced application monitors for services using the teaming link. and integrate this more complex link health/status logic into applications using libteam.