This paper investigates whether agents must learn causal models to generalize to new domains or if other inductive biases are sufficient. The authors show that any agent capable of satisfying a regret bound for a large set of distributional shifts must have learned an approximate causal model of the data generating process. This result has implications for several research areas, including transfer learning and causal inference. The paper begins by discussing the role of causal reasoning in human cognition and the challenges of developing agents that can adapt to new domains without explicitly learning causal models. It then presents a theoretical result showing that learning a causal model is necessary for robust adaptation. The authors define causal models and decision tasks, and discuss distributional shifts, which are changes in the environment or objective that agents must adapt to. The paper then presents three main results. The first shows that for almost all decision tasks, the underlying causal Bayesian network (CBN) can be reconstructed given optimal policies for a large set of domain shifts. The second result shows that for any regret bound, an approximate causal model can be learned from regret-bounded policies under domain shifts. The third result shows that having an approximate causal model is sufficient to identify regret-bounded policies. The authors interpret these results in terms of agents, transfer learning, and causal inference. They show that agents must learn causal models to adapt to known distributional shifts, and that transfer learning requires identifying causal relations between features and labels. The results also have implications for causal inference, showing that optimal policies under domain shifts can be used to identify all causal relations. The paper concludes by discussing the broader implications of these results for the development of robust and general AI. It suggests that causal models are fundamental to understanding and explaining the world, and that learning causal models is necessary for achieving strong robustness guarantees. The authors also discuss the practical challenges of learning causal models and the potential of using robust agents to elicit causal models from their policies.This paper investigates whether agents must learn causal models to generalize to new domains or if other inductive biases are sufficient. The authors show that any agent capable of satisfying a regret bound for a large set of distributional shifts must have learned an approximate causal model of the data generating process. This result has implications for several research areas, including transfer learning and causal inference. The paper begins by discussing the role of causal reasoning in human cognition and the challenges of developing agents that can adapt to new domains without explicitly learning causal models. It then presents a theoretical result showing that learning a causal model is necessary for robust adaptation. The authors define causal models and decision tasks, and discuss distributional shifts, which are changes in the environment or objective that agents must adapt to. The paper then presents three main results. The first shows that for almost all decision tasks, the underlying causal Bayesian network (CBN) can be reconstructed given optimal policies for a large set of domain shifts. The second result shows that for any regret bound, an approximate causal model can be learned from regret-bounded policies under domain shifts. The third result shows that having an approximate causal model is sufficient to identify regret-bounded policies. The authors interpret these results in terms of agents, transfer learning, and causal inference. They show that agents must learn causal models to adapt to known distributional shifts, and that transfer learning requires identifying causal relations between features and labels. The results also have implications for causal inference, showing that optimal policies under domain shifts can be used to identify all causal relations. The paper concludes by discussing the broader implications of these results for the development of robust and general AI. It suggests that causal models are fundamental to understanding and explaining the world, and that learning causal models is necessary for achieving strong robustness guarantees. The authors also discuss the practical challenges of learning causal models and the potential of using robust agents to elicit causal models from their policies.