Bridging the Gap: A Systematic Framework for Agentic AI Root Cause Analysis in Hybrid Distributed Systems

Abstract

The rapid expansion of microservices-based distributed systems across hybrid cloud and on-premises infrastructures has generated significant hurdles for root cause analysis (RCA) during production incidents. The microservices architecture has transformed the development and deployment of large-scale applications by dividing monolithic structures into smaller, stand-alone services. Studies show that it can take up to three hours to find the cause of a failure without automated tools. Enterprise cloud service providers such as AWS, Azure & GCP base their prices on guarantees of availability, which can be quite costly. This paper offers a thorough examination of the challenges associated with microservices and methodically contrasts traditional monolithic design with microservices methodologies. Recent modernization in agentic artificial intelligence holds significant potential for autonomous incident response. There are still big problems with scalability, explainability, causal inference, and support for hybrid infrastructure.

This paper presents AURORA (Autonomous Unified Root Cause Analysis through Observability and Reasoning Agents), a comprehensive multi-agent framework that integrates large language model-driven reasoning, hierarchical causal discovery via the Root Cause Discovery (RCD) algorithm, and federated observability for efficient root cause analysis in production environments. The main contributions are (1) a new theoretical framework for finding causes in the context of non-stationary interventions, with sample complexity bounds of O(n²log(n)); (2) a complete multi-agent architecture which combines the ReAct pattern with causal inference algorithms and achieving a top-5 recall of 94.3% on the Sock-shop benchmark, which is better than the leading RUN (91%) and RCD (89%) methods; (3) in-depth ablation studies that measure the contribution of each component across more than 500 failure scenarios involving systems with 15 to over 1000 services; (4) a taxonomy of failure modes that shows performance degradation; (5) a case study on production deployment involving 247 microservices, resulting in a 91% reduction in MTTR and a 74× ROI; and (6) comprehensive implementation guidance that makes it easy to adopt in the industry right away.

AURORA proves an absolute improvement of 3.3–5.3% over the best methods, as shown by a detailed empirical test on real distributed systems that was statistically significant (p < 0.01, Cohen's d = 0.85).