COBOL remains widely deployed, yet the lack of corpora that reflect real production code limits rigorous benchmarking of modernization techniques. We built an agentic AI pipeline that generates realistic COBOL programs. The pipeline first creates a code skeleton from a business specification, then applies iterative degradation that injects patterns and complexity targets extracted from actual production code. Degradation is guided by structural fidelity metrics such as function‑level depth, call‑graph density and data‑structure usage rate. Experiments on three programs from distinct business domains evaluate whether degradation reaches the target complexity while preserving business behavior. Results show the pipeline reliably produces syntactically valid code and moves structural complexity toward realistic levels. However, business behavior is not always retained, and optimizing structural metrics alone can produce programs that lack a plausible maintenance history. We discuss these limitations and outline a more realistic direction: generating legacy programs from scratch along a simulated development history.
Review