A study of nearly 150,000 real-world agent actions found AI increasingly handling low-risk IT tasks while human analysts iteratively shape and improve their actions.