During physical Human-Robot Interaction (pHRI) for limb mobilization, humans and robots may contribute simultaneously to motion. Some Assist-AsNeeded (AAN) strategies rely on models, while others are purely reactive. This paper presents a reactive AAN control for an end-effector robot in upper limb rehabilitation that weights commands dynamically based on local performance. Volunteers in tests followed a planar circular trajectory with visual feedback. Statistical analysis confirms that assistance is provided as needed, balancing performance across users and hands. Additionally, global metrics - including completion time, tracking errors, force, and disagreement - improve compared to standalone trajectories.