Studies on agreement production consistently report an increase in production errors in the presence of an attractor mismatching the agreement feature of the target. In contrast, results from comprehension studies are mixed, ranging from lack of effect to facilitation. We report two forced-choice experiments and two self-paced reading experiments on number and gender object-verb agreement in French in order to systematically explore the effect of a mismatching subject in the production and comprehension of object relatives. Results show that the presence of a mismatching subject penalizes sentence production, in line with reports of attraction, but consistently improves sentence comprehension in off-line comprehension measures, in line with similarity-based interference effects. We discuss the limits of classical models of sentence production and comprehension (Marking and Morphing and ACT-R), and favor a self-organizing sentence processing approach (SOSP), which accounts for both production and comprehension results through a single similarity-based mechanism of structure building.