We present a vision-guided agricultural automation framework with human-in-the-loop learning for harvesting, inspection, and occlusion removal. The system integrates lightweight 7-DoF manipulators with interchangeable soft grippers and dual-camera perception, mounted on cable/rail infrastructures for large-area operation. Three motion-generation strategies are implemented: YOLO-based visual servoing, diffusion-based policy, and a combined YOLO–diffusion policy. A teleoperation interface supports task-space demonstration, data correction, and motion intervention during deployment. The system is validated on raspberry harvesting and further demonstrated on bimanual coordination and human–robot collaboration tasks. The resulting framework integrates perception, human-in-the-loop control, and learned motion for large-scale agricultural manipulation. Note to Practitioners—Many agricultural robots are designed for a single task and rely on heavy platforms, expensive industrial arms, or task-specific control pipelines. In practice, farms also face changing crops, occlusions, uneven terrain, and situations that require human intervention. This paper presents a robotic framework that combines lightweight rail- or cable-mounted robotic arms, human-in-the-loop operation, and vision-guided motion learning. Operators can demonstrate and correct robot motions in real time, while learned policies handle repetitive or visually complex actions. The dual-arm teleoperation interface supports bimanual control and real-time collection of human motion data during operation. Different vision and control methods, such as object detection, learned motion policies, or their combination, can be selected based on the task. The system uses low-cost arms and modular hardware with minimal setup requirements. Demonstrations collected during operation can be used for policy training without expert programming or large datasets. The system was tested on harvesting, inspection, occlusion removal, and human–robot collaboration tasks. Current limitations include camera-based perception and the need for human intervention in unfamiliar situations. Additional sensors and improved perception methods can be integrated into the same framework. This allows the level of automation to increase gradually while retaining human supervision when needed.