Firehose

Researchers developed a novel agentic test-synthesis method, called the "Locksmith Loop," to validate legacy COBOL programs migrated to Java, achieving nearly complete coverage on two open-source programs and 91.90% branch coverage on a production-like COBOL program. The method uses iterative testing with mocks, mutations, and analysis to identify and fix bugs. The approach consistently improved coverage beyond input-search plateaus, demonstrating a deterministic validation method for agentic coding output. AI summary