SPEC-DRIVEN DEVELOPMENT FOR ENTERPRISE AGENTIC SOFTWARE ENGINEERING: A GOVERNANCE FRAMEWORK FOR RELIABLE AI-GENERATED SOFTWARE

AI coding assistants are evolving into agents that can plan, edit repositories, run tools, test software, and prepare deployable changes. Natural-language prompts are useful for expressing intent, but they rarely encode the complete set of enterprise requirements that govern architecture, cybersecurity, reliability, and operations. This study proposes the Specification Governance and Validation Framework (SGVF), a design-science artifact that treats structured specifications as an external control plane for agentic software engineering. SGVF separates requirements into functional, architecture, security, and operational contracts; introduces a Specification Compliance Score (SCS) with mandatory control vetoes; classifies change risk to determine permissible autonomy; and records traceability evidence from business intent through deployment. The framework was evaluated analytically through six enterprise scenarios and qualitative comparison with traditional, prompt-based, vibe-coding, and ungoverned agentic approaches. The evaluation shows how SGVF produces explicit release decisions for authorization failures, architecture drift, vulnerable dependencies, operational regressions, privileged changes, and compliant low-risk work. The study does not claim measured productivity or defect-reduction gains. It provides a reproducible governance model and a basis for future empirical validation of specification-driven enterprise AI development.