Delineating agricultural field boundaries is key to building sustainable supply chains, as boundaries link agricultural products directly to their sites of production and serve as informative units of analysis for land use change. However, manually collecting boundary data from the ground or from high-resolution satellite images is costly and slow, limiting its application at scale. This technical note describes Trazo, an iterative framework that addresses this problem by incorporating new expert-annotated field boundary data into deep learning segmentation models for automated delineation from satellite imagery. We developed the method using high-quality, diverse training data from 17 soy-producing ecoregions in South America together with seasonal Sentinel-2 imagery, producing a suite of tailored segmentation models, called Trazo, and a pipeline that deploys them at scale. The first release maps nearly 11 million field boundaries across South America. The aim is to provide open, reliable plot-level data and tools that help supply chain actors strengthen traceability, assess where deforestation and ecosystem conversion are occurring within their supply chains, and comply with sustainability commitments and regulations such as the European Union Deforestation Regulation (EUDR).