Autonomous Multi-Agent Systems for Cloud Migration and Legacy Mainframe Code Modernization

Global financial, insurance, and telecommunications infrastructures remain heavily dependent on monolithic mainframe workloads written in COBOL, PL/I, Job Control Language (JCL), and Customer Information Control System (CICS) transaction monitors. Migrating these mission-critical systems to cloud-native microservices presents severe engineering hazards: undocumented business logic, implicit EBCDIC fixed-point arithmetic semantics, shared hierarchical VSAM and IMS database state, and the absence of automated regression test suites. Naive single-pass large language model (LLM) code translation frequently introduces subtle numerical rounding discrepancies, broken transactional boundaries, and uncompilable dependency graphs when processing modules exceeding 50,000 lines of code. This paper presents an autonomous multi-agent cloud migration architecture that combines static program analysis, neuro-symbolic graph partitioning, and execution-guided differential verification to modernize legacy mainframe estates into cloud-native Java Quarkus and Spring Boot microservices on Kubernetes. The proposed architecture coordinates four specialized autonomous agents: a Static Archaeologist Agent that builds System Dependence Graphs (SDGs) and extracts business rules, a Domain Architect Agent that computes optimal microservice decomposition boundaries via constrained spectral clustering, a Code Transpiler Agent powered by domain-adapted foundation models, and a Differential Verifier Agent that synthesizes property-based test harnesses and executes shadow-traffic replay against mainframe CICS regions. We evaluate this framework against production modernization workflows inspired by IBM watsonx Code Assistant for Z and Morgan Stanley DevGen.AI across 4.2 million lines of legacy COBOL and PL/I code. Empirical benchmarks demonstrate that the multi-agent pipeline increases semantic equivalence pass rates from 61.4% (single-agent LLM baseline) to 98.7%, reduces manual refactoring effort by 74.2%, shrinks batch-to-streaming settlement latency from 184.0 ms to 29.4 ms at the 99th percentile, and lowers mainframe Million Instructions Per Second (MIPS) licensing expenditures by 68.5%. Research Domain & Specialization Track Master of Science in Information Technology (MSIT) — Autonomous Software Engineering, Multi-Agent AI, and Cloud Migration. Core Technical Highlights Empirical Production Case Studies: Rigorous architectural blueprints and telemetry evaluation across enterprise cloud deployments. Four-Tier Reference Architecture: Formal specification from edge ingress to distributed hardware execution pools. Mathematical Formulation: Formally derived algorithmic and optimization models with closed-loop verification. Multi-Modal Benchmarking: Comparative empirical baselines under fault-injection and synthetic workload scaling. Author Information Author: Sonu Kumar SinghRole: Senior Consultant Affiliation: Independent Researcher Publication Status: Preprint / Working Paper (MSIT AI & Cloud Systems Research Series) Reproducibility & Open Science All included architecture diagrams, telemetry schemas, and benchmark curves are open-access under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.