We summarize REA (Routed Embodied Agent Runtime), a unified framework for asynchronous coordination between deliberative System 2 foundation models and reactive System 0 and System 1 policies. REA resolves the incompatibility between slow, high-level reasoning and fast, low-level motor control by introducing three core components: a route memory storing pre-measured policy success rates, a boundary-triggered judge gate that invokes the brain only upon execution boundary violations, and a shared-budget escalation ladder enabling bounded recovery without blocking fast pathways. Critically, REA enforces a golden rule that no running policy waits on the brain, eliminating the stall-and-fallback cycle observed in prior agentic loops. Evaluated on 3,000 LIBERO-Pro episodes across 60 tasks, REA achieves a 95.9% success rate, surpassing the best single-agent baseline (94.7%) while reducing control steps by 25% and control time by 17%. In a 100-episode RoboTwin 2.0 study, REA attains 84% success with VLA policies and 86% with VLA+code, outperforming the strongest single VLA baseline MEDex at 80%. The runtime further transfers to BEHAVIOR-1K and a real Unitree G1 humanoid, achieving the highest observed success among compared methods, demonstrating that learned and programmatic policies can operate synergistically within a single asynchronous control stack.