Coding agents hit the one obstacle they were told to avoid, and the fault is planning.

arXiv 2609.20822 pairs each robot manipulation task with a forbidden obstacle and finds the agent collides with it in most cases. The diagnosis generalizes well past robotics: the agent reasons about the obstacle in its traces and the prompt forbids touching it, so neither perception nor instruction failed. The stated constraint simply never became a planning priority. SafeHarness grounds objects as bounding boxes, draws candidate waypoint routes, plans-verifies-replans before executing, and reaches 71.9% task success and 87.5% collision avoidance against the same agent's 31% and 58%. arXiv