This study was designed to (1) investigate the procedure responsible for successful written subject-verb agreement, and (2) describe how it develops across grades. Students in Grades 3, 5 and 12 were asked to read noun-noun-verb sentences aloud (e.g., "Le chien des voisins mange" ["The dog of the neighbors eats"]) and write out the verb inflections. Some of the nouns differed in number, thus inducing attraction errors. Results showed that third graders were successful because they implemented a declarative procedure requiring regressive fixations on the subject noun while writing out the inflection. A dual-step procedure (Hupet, Schelstraete, Demaeght, & Fayol, 1996) emerged in Grade 5, and was fully efficient by Grade 12. This procedure, which couples an automatized agreement rule with a monitoring process operated within working memory (without the need for regressive fixations), was found to trigger a mismatch asymmetry (singular-plural > plural-singular) in Grade 5. The time course of written subject-verb agreement, the origin of agreement errors and differences between the spoken and written modalities are discussed.