Claude Fable 5 programs UR5e arm using force sensors

The author argues that agentic robotics flips the traditional build order by placing high-level frontier AI planners first and shrinking low-level motor controllers. Recent demonstrations show coding agents successfully writing their own functional Python controllers for robot arms using tool calls and sensor feedback, often zero-shotting benchmarks without robot-specific fine-tuning. The author predicts that agent-written programs using touch or force will match learned policies on insertion tasks and that agent planners will dominate benchmarks by mid-2027. Open article →