AutoSynthData: Generating Training Data for Enterprise Agents

AutoSynthData is a system that generates training data for enterprise agents by identifying what a model struggles with and creating new tasks to teach those capabilities. Enterprises need agents tailored to their specific environments, systems, and workflows, but turning individual model failures into useful training data requires many realistic tasks that the agent cannot yet solve consistently. The system works by evaluating a target model's performance, using a stronger teacher model to understand what could be learned, and then automatically generating varied tasks with verifiers to check if solutions are correct. This approach creates a continuously improving curriculum where the system identifies remaining capability gaps and generates new training tasks accordingly.