In this paper, I present evidence for variable agreement with anaphors in Tatar. I show that inflected reflexives trigger co-varying person agreement as DP/nominalization subjects and as complements of postpositions, which appears to contradict the generalization on the anaphor agreement effect (AAE). At the same time, inflected reciprocals induce 3p agreement on external targets. These data are puzzling in two aspects. First, it is unclear how to derive co-varying agreement with inflected reflexives because it cannot be handled as a regular exception to AAE predicted to arise by the agreement-based theory if the antecedent of the anaphor is positioned lower than the agreement target. Secondly, the difference between reflexives and reciprocals with respect to external agreement looks enigmatic. I propose that Tatar reflexives and reciprocals, despite their superficial resemblance, have different internal structures, which in turn bring about differences in their feature sets, and external agreement reveals these differences. As to AAE violations, I propose that the Tatar data can be accounted for under the feature sharing approach whereby the features on the anaphor and on the external probe are first identified as instances of the same feature set and then valued by the anaphor’s binder.