This paper applies a feature geometry for person and number agreement to developmental patterns in second language acquisition. It was predicted that the more highly-specified dependent nodes (1st, plural, finite) would successfully emerge later than their underspecified counterparts (3rd, singular, nonfinite). A corpus of interview data from two proficiency levels, including over 900 lexical main verbs, was analyzed for accuracy and error type. For person and finiteness, the prediction was supported: accuracy in 1st and finite emerged later than 3rd and nonfinite. Plural agreement remained somewhat problematic, even among higher-proficiency learners. For person, number, and finiteness, underspecified forms were favored as defaults, indicating systematic patterns in both variation and development.