The Limits of Predicting Agents from Behaviour