The real world is too complex to model accurately. Autonomous agents and robots that perform complex tasks in the real world, ranging from handling inventory in warehouses to driving, will inevitably encounter scenarios that are not fully described in their symbolic models used for decision-making. To handle such unexpected scenarios, agents often rely on human assistance to complete the task, restore safety, or refine the model. While these interventions can restore safety in the short term, they are reactive, require extensive human effort, and fail to generalize. Consequently, translating empirical success from structured environments to real-world settings remains challenging and hinders long-term autonomy. Just as human decision-makers adapt by reflecting on their limitations and seeking information or assistance, as needed, reliable long-term autonomy requires autonomous systems to proactively recognize their model limitations, seek additional information to refine their models, produce context-appropriate behaviors, and recover from unexpected runtime errors, with minimal reliance on humans. This project will develop an introspective reasoning paradigm for autonomous agents, grounded in three core properties: (1) model cognizance—autonomous and proactive identification and repair of model limitations, through limited user interactions; (2) contextual planning—prioritizing context-appropriate objectives and balancing tradeoffs in multi-objective planning, by determining ``what matters when''; and (3) resilient execution—detect and recover from unforeseen execution-time errors with minimal disruption to task performance, via self-monitoring and online replanning. By closing the loop between self-monitoring and behavior adaptation, this project lays the foundation for autonomous systems that operate reliably in unstructured environments, while gradually reducing the human effort involved. This award reflects NSF's statutory mission and has been deemed wor