This paper addresses advanced agent-based inference models operating under resource constraints and the potential for failure. These models interact with humans and solve sequential decision-making problems based on (approximate) utility functions and internal models. In resource-constrained or failure-constrained problems where resource exhaustion can force the termination of action sequences, agents face implicit trade-offs in reconfiguring utility-driven rational behavior. Furthermore, because these agents typically act at the behest of humans, asymmetries in constraint exposure can lead to unexpected misalignments between human goals and agent incentives. This paper formalizes these settings using a survival bandit framework, presents theoretical and experimental results quantifying the impact of survival-driven preference shifts, identifies conditions under which misalignment occurs, and proposes mechanisms to mitigate the occurrence of risk-seeking or risk-averse behavior. Consequently, this study aims to enhance our understanding and interpretability of the emergent behavior of AI agents operating under such survival pressures and to provide guidance for the safe deployment of such systems in critical resource-constrained environments.
