Evaluating the Robustness of Collaborative Agents