Rising interest in artificial intelligence in education reinforces the demand for evidence-based implementation. This study investigates how tutor agents' physical embodiment and anthropomorphism (student-reported sociability, animacy, agency, and disturbance) relate to affective (on-task enjoyment) and cognitive (task performance) learning within an intelligent tutoring system (ITS). Data from 56 students (M = 17.75 years, SD = 2.63 years; 30.4% female), working with an emotionally-adaptive version of the ITS "Betty's Brain", were analyzed. The ITS' agents were either depicted as on-screen robots (condition A) or as both on-screen avatars and physical robots (condition B). Physical presence of the tutor agent was not significantly related to task performance or anthropomorphism, but to higher initial on-task enjoyment. Student-reported disturbance was negatively related to initial on-task enjoyment, and student-reported sociability was negatively related to task performance. While physical robots may increase initial on-task enjoyment, students' perception of certain characteristics may hinder learning, providing implications for designing social robots for education.