Layered control is a concept in decision-making that separates responsibilities across different optimization problems designed to collectively perform an overall task. This concept is well-understood among robotics and control engineers, since computation is a fundamental bottleneck in real-time, safety-critical applications. However, the failure of a system performing a task is often attributed to the layered approach itself. In this thesis, we claim that the lack of principled design of such layers and an understanding of their interplay is the underlying cause. We demonstrate this claim across four problems. For canonical linear-quadratic regulation under tight computational budgets, we present a decomposition method that splits each iteration into a dynamics-feasibility step and a constraint-feasibility step. Warm-starting from previous iterates with early stopping guarantees closed-loop stability and recursive feasibility. We then extend principled layered design beyond model-based methods, applying an optimization decomposition to derive a data-driven trajectory generator that succeeds where existing layered designs fail to plan feasible trajectories. For the longstanding challenge of dynamic obstacle avoidance, we outperform existing layered methods and expose the brittleness of reactive and end-to-end approaches. Finally, we formulate the novel problem of impact-aware safety in contact-rich tasks and solve it with a layered approach that balances task performance and safety. Throughout, we emphasize that layers do not by themselves cause failures, rather the lack of safety-aware design across layers is the root cause. This provides a foundation for designing tightly integrated perception-action loops and layered communication for multi-agent systems.