Abstract Educational software based on teachable agents has repeatedly proven to have positive effects on students’ learning outcomes. The strongest effects have been shown for low-performers. A number of mechanisms have been proposed to explore this outcome, in particular mechanisms that involve attributions of social agency to teachable agents. Our study examined whether an expression of high versus low self-efficacy in a teachable agent would affect low-performing students with respect to their learning outcomes and with respect to a potential change in their own self-efficacy. The learning domain was mathematics, specifically the base-ten system. Results were that the learning outcomes of low-performers who taught a low self-efficacy agent were significantly better than the learning outcomes of low-performers who taught a high self-efficacy agent. There were no effects from the manipulation of self-efficacy expressed by the teachable agent on changes of the low-performing students’ own self-efficacy.