Adyen and Hugging Face release DABstep data agent benchmark for multi-step reasoning

AI Impact Summary

DABstep is a benchmark co-developed by Adyen and Hugging Face to evaluate agent-based, multi-step data analysis on real-world tasks. It includes 450+ tasks that mix structured and unstructured data and require sequential reasoning plus code execution, reflecting enterprise workloads. Early results show the most capable reasoning-based agents achieving only 16% accuracy, highlighting a substantial gap between current model capabilities and practical data-analysis workflows. This indicates enterprise teams should plan for significant tooling and integration work to close the gap, including robust data access, verification, and iterative reasoning pipelines when deploying autonomous data agents.