Trace learning for self-improving agents