AMADRS: An Autonomous Multi-AgentDisaster Recovery System : Design and Evaluation of AI-Driven Recovery for Hybrid Cloud Environments

Sammanfattning

Modern cloud systems operate as distributed and interdependent architectures in which localized failures can rapidly escalate into service-wide disruptions. Existing disaster recovery approaches remain largely manual or script-driven, introducing delays in failure interpretation, decision-making, and coordinated remediation, particularly in multi-component and multi-region environments. This thesis presents an Autonomous Multi-Agent Disaster Recovery System (AMADRS) that integrates AI-assisted reasoning with Infrastructure-as-Code (IaC) execution to enable closed-loop recovery workflows in cloud environments. The system is designed as acoordinated set of agents responsible for failure detection, contextual analysis, recovery orchestration, Infrastructure-as-Code generation, and automated remediation. A large language model is integrated as the decision-support component responsible for interpreting infrastructure state, selecting recovery strategies, and coordinating recovery actions. The artifact was evaluated using a Design Science Research methodology through controlled fault-injection experiments across four recovery scenarios, including same-region recovery, full-stack recovery on the same-region, cross-region failover, and cross-cloud recovery. Recovery performance was evaluated using Mean Time to Recovery (MTTR) and compared against a human-led recovery baseline under equivalent experimental conditions. The evaluation results show that the proposed system achieved lower MTTR across all evaluated scenarios, with the largest improvements observed in cross-region and cross-cloud recovery cases involving higher coordination complexity and dependency management overhead. In simpler same-region recovery scenarios, the performance differences between AI-assisted and human-led recovery were smaller, although the autonomous system still achieved lower measured recovery times. The findings further indicate that automation primarily reduces coordination and orchestration overhead, while infrastructure provisioning latency remains constrained by cloud-provider managed service initialization times. This research demonstrates the feasibility of integrating AI-assisted reasoning into operational recovery workflows and provides empirical evidence that autonomous multi agent coordination can improve recovery efficiency in distributed cloud environments. The proposed approach establishes a foundation for future autonomous resilience and recovery systems in hybrid and multi-cloud infrastructures.