I think this paper is nevertheless of interest because it describes a large number of cases in which the kinds of syntactic treatments of gender agreement that linguists have assumed in the past are all inadequate; in those cases, semantics and pragmatics are the only possible satisfactory sources of the observed generalizations about agreement. The correct conclusion to draw from the evidence as a whole, I would now maintain, is that a complete description of languages (those like English anyway) must include BOTH a semantic AND a syntactic mechanism for gender and number agreement.
(I would note that other papers of mine argue that the only really adequate description of certain linguistic constructions requires the recognition that two different analyses of them are necessary, simultaneously. These include both a syntactically- governed and a pragmatic analysis of control of some VP complements ('Toward a non-grammatical account of thematic roles', with Bill Ladusaw), and analyses of some constituents as both adjuncts and complements at the same time ('The Dual Analysis of Adjuncts/Complements in Categorial Grammar', 2003). Morpho-syntactic as well as and semantic-pragmatic analyses of agreement should be added to the list.)
The argument in the paper that is in effect "circular"---actually, trivializing---is that some difficult cases for the natural-gender analysis can be dismissed because it is a relevant property of an individual or class of individuals that it/they are standardly referred to by a noun of a certain gender and number, and that that kind of property satisfies the 'semantic' agreement hypothesis for those cases.
Now, it undoubtedly is correct that this kind of property is one that ultimately has to be included in a semantics of natural language, in order to account for the truth conditions of a variety of sentences, one of which is Quine's example "Giorgione is so-called because of his size" (cited by Quine for a different reason), or "Delicia has a pretty name", or other (more interesting) cases. But if this observation were viewed as a compelling argument that agreement is semantic in nature, that would render the whole question of syntactic-vs-semantic agreement trivial (as that question is normally understood by linguists, including the authors of the paper), because it would apply to every conceivable case of agreement and thus rescue any imaginable counterexample to the semantic agreement hypothesis. But the Giorgione-example and its cousins all involve meta-linguistic reference, something that has many kinds of manifestations in natural language, and such sentences have semantic quirks that other natural language sentences do not have. Appealing to all those properties in linguistic argumentation about the semantic analysis of sentences in general would wreak havoc for many widely-accepted principles of linguistics, so isolating those properties from regular linguistic argument, at least for the present, is surely advisable. Hence a syntactic or a quasi-syntactic, quasi-semantic-pragmatic analysis of some kind is presumably necessary for at least certain cases of `long-distance' agreement.
A copy of the paper (scan of the published version) is available here:
[with Pauline Jacobson] "Agreement as a Semantic Phenomenon" (.pdf) 1988. Proceedings of Eastern States Conference on Linguistics, Columbus: OSU Dept. of Linguistics.