SIGMMA: Skill-Based Multi-Agent Generation of Test Cases from Multimodal User Manuals

SIGMMA generates test cases from use cases in two phases. Phase 1 extracts use cases page by page from a PDF user manual with a vision-language model. Phase 2 generates one test case per use case and refines it in a loop between a generator agent and a validator agent until the test case satisfies four criteria: completeness, atomicity, clarity, and traceability. The domain rules of each agent are packaged as SKILL.md files following the Agent Skills standard.