An enduring question in education is whether team-based peer learning methods help improve learning outcomes among students. We randomly assign around 10,000 middle school students in Karnataka, India, to alternative peer learning treatments in Math and English that vary the intensity of collaboration. Teamwork with co-coaching outperforms simple teamwork and incentive treatments by increasing the test scores by about 0.25 standard deviation, but only in Math. This is both statistically and economically significant for students at the bottom of the ability distribution. We develop theoretical conditions under which teamwork with co-coaching outperforms simple teamwork as a peer-learning method.