Literature Survey: Legacy Modernization with Multi-Agent LLMs and Behavioral Specifications

Overview

This survey examines research related to legacy system modernization and identifies key challenges that remain unaddressed by existing work.

Established Research Areas

1. LLM-Based Legacy Modernization

[citation_1l9tk0] Generative AI enables transformation of monolithic architectures into modern microservices and facilitates code translation, documentation generation, test case creation, and intelligent refactoring recommendations. [citation_1l9tk4] Structured frameworks combined with human review processes can modernize legacy applications in contemporary languages like Java and .NET.

However, [citation_1l9txk] large language models face significant context limitations—even one-million-token context windows are insufficient for the largest legacy codebases, and tools are largely file- and session-scoped, with understanding lost when sessions end.

2. Multi-Agent Systems for Software Engineering

[citation_1l9xlz] Software development can require cooperation among multiple members with diverse skills, motivating exploration of agentic frameworks. [citation_1l9x0q] A team-based approach models software engineering as a collaborative process with agents having distinct roles and execution environments. [citation_1l9wvk] In practice, multi-agent coordination for debugging workflows achieved a 93% reduction in time-to-root-cause.

3. Program Comprehension and Code Understanding

[citation_1l9ur2] Program comprehension involves top-down and bottom-up processes. [citation_1l9uva] In the era of generative AI, the critical skill has shifted toward understanding and maintaining AI-generated code. [citation_1l9umn] Code understanding challenges include navigation and question-answering.

4. Business Logic Extraction from Legacy Systems

[citation_1l9ysv] Business rules in legacy systems are rarely documented—they live in control flow patterns, conditional branches, exception handlers, and configuration files. [citation_1l9y9m] Semantic business rule extraction seeks to infer the original intention of procedural logic, complicated by intertwining of technical and business logic. [citation_1l9yl3] Model-driven reverse engineering frameworks extract business logic from legacy COBOL systems.

5. Program Transformation and Semantic Refactoring

[citation_1l9yy1] Syntax-driven refactoring focuses on syntactic transformations, while semantics-driven refactoring uses formal verification to reason over program behavior. [citation_1l9yxm] Program synthesis can automate refactoring decisions by verifying observational equivalence between original and refactored code.

Critical Research Gaps

Gap 1: Session Persistence and System-Wide Understanding

[citation_1l9txk] AI tools operate in session-scoped manner, with understanding lost when sessions end.

Gap 2: Behavioral Correctness During Transformation

[citation_1l9ysv] A syntactically correct translation can pass compilation but silently break edge-case handling, alter validation logic, or drop constraints that were never written down.

Gap 3: Behavioral Specification Extraction

[citation_1l9ysv] Silent behavioral changes during migration remain a central risk when business rules are not documented.

Gap 4: Large-Scale Codebase Context

[citation_1l9txk] Even one-million-token context windows are insufficient for the largest legacy systems spanning hundreds of thousands of files.

Gap 5: API Integration During Code Translation

[citation_1l9tka] Direct application of LLMs for code translation often produces syntactically valid outputs that fail to compile or integrate within existing production frameworks.

Gap 6: Intermediate Representations for Modernization

[[citation_1l9uvs] Program transformation literature discusses representations ranging from source code to intermediate representations to specification.

Conclusion

The surveyed literature shows mature individual techniques for legacy modernization, multi-agent orchestration, program understanding, business logic extraction, and semantic refactoring. Significant challenges remain in integrating these capabilities into systems that maintain behavioral correctness while scaling to large codebases.