New LLM-orchestrated multi-agent framework automates full BDaaS lifecycle

A new paper on arXiv (2606.17915) presents a trustworthy self-composable BDaaS framework built on LLM-orchestrated multi-agent collaboration. The system decomposes the entire data science lifecycle into eight specialized agents: data ingestion, cleaning, feature engineering, AutoML training, model evaluation, MLOps deployment, monitoring, and drift detection. A central LLM orchestration layer coordinates these agents, validates intermediate outputs, manages workflow context, and allows dynamic composition of pipelines. The framework includes shared artifact governance, reproducibility support, human-in-the-loop checkpoints, and drift-aware feedback loops.

Evaluation uses controlled tabular benchmark datasets with missing values, categorical variables, outliers, class imbalance, and simulated covariate drift. Compared to manual ML, AutoML-only, and single-agent LLM baselines, the multi-agent system achieves competitive predictive performance while significantly improving lifecycle-level reliability—including workflow completion, artifact traceability, deployment readiness, reproducibility, and drift recovery. The paper spans 7 pages with 3 figures and 5 tables, suggesting that LLM-orchestrated multi-agent systems can extend conventional AutoML toward production-oriented, adaptive BDaaS automation.