When enterprises modernize legacy systems, the hard part is not swapping the technology stack — it is preserving the business logic buried inside decades-old code that nobody documented but everything depends on. Most LLM approaches treat this as translation: feed in COBOL, get out Python. The syntax converts fine. The behavior breaks silently. AgentModernize treats modernization as a behavior-preservation problem. Four specialized LLM agents extract business rules, organize them into a Behavioral Specification Graph (BSG), generate modern code under behavioral contracts, and validate equivalence with a feedback loop. The BSG acts as a trust boundary — every extracted rule is inspectable before any code gets written. On LegacyModernize-8 (eight scenarios, telecom + banking), AgentModernize with GPT-4o-mini is the only method to pass any gold-standard behavioral tests. All single-prompt and chain-of-thought baselines score 0.0%. Remove the feedback loop, and the pipeline also scores zero. With GPT-5.3-codex, a single prompt beats the full pipeline in raw score (20.4% vs 11.0%) — but produces no traceable artifacts. No rule inventory, no BSG, no equivalence report. In regulated domains, that audit trail matters as much as the code. Legacy modernization is not code translation. It is contract-driven behavioral recreation, where preserving semantics and maintaining auditability are as important as generating working code.
Sheikh Nazib Ahmed (Sun,) studied this question.