A dataset of real-world coding agent sessions from open-source developers is presented, providing an empirical characterization of agent usage and failure modes. The dataset shows that agents remain inefficient in natural settings and introduce more security vulnerabilities than human-authored code.