Agentic Robotics (AR) represents a paradigm shift in robot intelligence development, using multi-agent AI systems to generate and iteratively improve robot programs offline rather than relying on demonstration data or manual engineering. Researchers at UC Berkeley and NVIDIA developed Graph-as-Policy (GaP), a graph-based framework that enables coding agents to compose modular robot skills into interpretable control systems, achieving improved performance on benchmarks while addressing simulation bottlenecks through inverse physics approaches.