Recent work has explored how complementary strengths of humans and artificial intelligence (AI) systems might be productively combined. However, successful forms of human-AI partnership have rarely been demonstrated in real-world settings. We present the iterative design and evaluation of Lumilo, smart glasses that help teachers help their students in AI-supported classrooms by presenting real-time analytics about students' learning, metacognition, and behavior. Results from a field study conducted in K-12 classrooms indicate that students learn more when teachers and AI tutors work together during class. We discuss implications for the design of human-AI partnerships, arguing for participatory approaches to research in this area, and for principled approaches to studying human-AI decision-making in real-world contexts.