Eco-FinOps: A Causal-Agentic Framework for Energy-Efficient and Explainable Cloud Cost Optimization

Abstract

Two converging crises—uncontrolled operational expenditure and significant environmental hazards—have been aggravated by the exponential scaling of cloud infrastructure. This has led to a greater focus on Financial Opera-tions (FinOps) efficiency. Current monitoring methods exhibit a major flaw: passive dashboards require constant human monitoring and generate alert fatigue, while deep learning models require heavy GPU computing power and act as opaque "black boxes." This lack of interpretability contradicts the sustainability goals these tools promise to support. This study pro-poses Eco-FinOps, a causal-agentic framework capable of optimizing cloud costs autonomously and energy-efficiently, designed strictly for commodity CPU environments without discrete GPU requirements. The system uses Polars-based telemetry ingestion, the PC algorithm for causal graph construction, and DoWhy for effect estimation to mathematically distinguish genuine workload fluctuations from inefficient resource usage. This study tar-gets the "zombie" resource pattern—provisioned instances utilizing merely 0.2-5% CPU while maintaining 60-90% memory consumption. Furthermore, a LangGraph-coordinated Retrieval-Augmented Generation (RAG) pipeline enables Small Language Models to generate auditable, human-verifiable remediation scripts. Evaluated on five fault-injection scenarios from the Al-ibaba Cluster Trace dataset, the framework achieves a Precision of 0.89, Recall of 0.86, and F1-score of 0.87. Causal verification successfully elim-inates 39% of false positive anomalies, raising Precision by 38 points over static threshold rules and outperforming neural autoencoder baselines by 15 F1 points.