The farther away one gets from a chunk of language that has some agreement to come from a later part of an utterance, the less likely one is to recall what that agreement is supposed to be, and the less likely one is to notice if it's not present. That's what happens with all kinds of agreement disparities, like certain tag questions.
- Your family is still in Ohio, aren't they?
- No one can see us here, can they?
- Anybody can do that, can't they?
Basically, we only parse as much as we need to, and the memory of earlier parses gets wiped by the parsing of later parts. Since most of English agreement phenomena contribute nothing to meaning, they're dispensible.