Agentic Architecture for Data-Centric AutoML

Data-Centric Artificial Intelligence (AI) focuses on improving data quality and curation as the main driver of Machine Learning (ML) performance. Although Automated Machine Learning (AutoML) has reduced the cost of model selection and hyper-parameter tuning, most systems remain model-centric and poorly aligned with Machine Learning Operations (MLOps) requirements such as controllability, traceability, and process-level explainability: automated choices about preprocessing, pipeline structure, and validation are often opaque and difficult to revise, despite their strong influence on model behaviour. We propose AGE-ML, a specification-driven, agentic architecture for Data-Centric AI. AGE-ML is data-centric in the sense that pipeline construction is conditioned on the data at hand: the selected workflow may include preprocessing, feature handling, rebalancing, and other data-preparation steps required by the dataset properties and by the selected estimators. A human-readable specification language enables data scientists to encode assumptions, constraints, evaluation criteria, and design preferences across the pipeline. From the specification, a search space for pipeline generation is algorithmically constructed and explored to support controlled automation and reproducible execution traces. Then, Large Language Models (LLMs) operate within a controlled agent loop to interpret specifications, generate and validate pipelines, and produce structured explanations for them. Results show improved traceability of pipeline generation, iterative pipeline refinement, and stronger integration with MLOps practices such as experiment tracking and reproducibility. Finally, we discuss limitations and future directions for AGE-ML.