SafeHarness: coding agents collide with the one obstacle they were told to avoid, and the fault is in planning, not perception

arXiv·medium signalXBlueskyLinkedInCopy link

arXiv 2609.20822 (2026-09-17) pairs each robot manipulation task with an obstacle the robot must not touch, and finds the coding agent hits it in most cases while treating task completion as its only objective. The diagnosis is precise and generalizes past robotics: the agent reasons about the obstacle in its traces and the prompt forbids touching it, so neither perception nor instruction fails, the stated constraint simply never becomes a planning priority. Splitting manipulation into a route phase and a contact moment shows the model has no notion of a clearing route and no replanning when a route becomes infeasible. SafeHarness grounds objects as bounding boxes, draws candidate waypoint routes, plans-verifies-replans before executing, and picks contact positions that avoid the obstacle, reaching 71.9% task success and 87.5% collision avoidance against the same agent's 31% and 58% without harnesses.