arXiv paper: AtumAI: A Principled Framework for Agentic Generation of Datacenter Control-Plane Policies

A new arXiv paper introduces AtumAI, a framework for using agentic AI to generate datacenter control-plane policies. It converts plain-language goals into formal, machine-checkable specifications covering objectives, constraints, decision variables, and evaluation methods, then uses an evolutionary design loop combining a diffusion model, evolutionary algorithm, and surrogate model to propose, test, and refine policies. The authors say this addresses limitations of off-the-shelf agentic AI, including weak handling of formal constraints, lack of transfer across tasks, and narrow exploration. In evaluations covering workload placement, resource scaling, and power management, the generated policies consistently outperformed expert-engineered baselines, according to the paper. The authors also claim the framework could reduce the work needed to adapt to a new task from months of engineering to writing a task description.