It's not clear to me what is causing this, but it just seems to naturally happen over time. Agents seem to be left in offline states but are not deleted at the end of their workload.

We periodically run doDoDelete() through the Script Console to delete these :-/