In a world filled with a multitude of complex and varied tasks, how can we empower an agent to accomplish tasks it has never encountered during training? Recent research endeavors to address this by employing a high-level planner to orchestrate a novel task as the composition of trained primitive skills, which can be executed by low-level controllers step by step. We formulate this method as Composable Generalization Agents (CGAs). Despite the promising future, the community is not yet adequately prepared for CGAs, particularly due to the lack of primitive-level datasets. In this paper, we propose a primitive-level real-world robotic dataset, namely RH20T-P, which contains about 33000 video clips covering 44 diverse and complicated robotic tasks. Each clip is manually annotated according to a set of meticulously designed primitive skills, facilitating the future development of CGAs.  To validate the effectiveness of RH20T-P, we also construct a potential and scalable agent based on RH20T-P, called RA-P. Equipped with two planners specialized in task decomposition and motion planning, RA-P has excellent spatial perception and can adapt to novel physical skills through composable generalization.