This paper proposes a multi-agent approach to overcome the limitations of conventional linear auditory scene analysis (ASA) systems. Conventional ASA systems sequentially process sound source localization, segmentation, and classification, resulting in long response times and significant impact on subsequent stages due to errors in early stages. The proposed multi-agent auditory scene analysis (MASA) system performs localization, segmentation, and classification tasks in parallel and compensates for errors through a reciprocal feedback loop. For example, the quality of the separation results is used to correct localization errors, and the classification results are used to reduce the sensitivity of localization to interference. This makes MASA robust to local errors and provides fast response times without increasing complexity. The proposed MASA system is provided as an open framework using JACK (Acoustic Acquisition and Reproduction) and ROS2 (Inter-Agent Communication), allowing easy user agent integration.