In Spanish clitic-doubling constructions, the clitic should agree in number with its coreferential doubled noun phrase. However, the present corpus analysis with data from 21 Spanish varieties reveals that, under certain structural configurations, number agreement is not always realized on the third-person dative clitic. In fact, the data shows that non-agreement appears to be the norm when the indirect object is a lexical noun phrase (77 vs. 23%). In this paper, I investigate two possible explanations for this phenomenon: (i) a processing account via an attraction effect and (ii) a syntactic account based on intervention effects. These two hypotheses make clear and testable predictions that I examine by means of conditional inference trees and Bayesian generalized mixed-effects logistic regression modeling. The results of the statistical analyses are incompatible with an intervention account because this type of phenomenon is not sensitive to semantic features of the intervening element or to the true controller of agreement. Thus, I propose that the data is best analyzed as the interplay between attraction and the morphosyntax of the unmarked. In Spanish, this results in attraction effects from the DO in the unmarked word order and inanimate IOs showing a sort of differential dative marking, where animate IOs show a preference for full agreement. The findings reported herein show evidence of a complex and highly dynamic agreement mechanism of the clitic and highlight the probabilistic nature of morphosyntactic processes.