Doug Rosenberg, Parallel Agile, Inc. and Caltech CTME, with Brian Moberley and Dr. Rick Hefner.
The emergence of Large Language Model (LLM) agents and "Agentic AI" has introduced a paradigm shift in software development, often referred to as "VibeCoding." Systems such as Claude Code and OpenAI Codex can generate complex codebases from high-level prompts, but they lack the formal constraints required for mission-critical systems. The result is a consistency gap: software outputs that are highly dependent on prompt engineering, difficult to reproduce, and nearly impossible to verify against system-level requirements.
This presentation proposes a bridge between formal systems engineering and agentic software execution. Building on our established framework for AI-Assisted MBSE (AIM), which automates the generation of SysML v2 models for complex hardware-software systems, we extend these techniques to the software domain. Instead of feeding agents informal natural language, we utilize AIM code-generation templates to produce structured, machine-readable behavior specifications. We demonstrate this approach through a comprehensive case study, illustrating how formal specifications reduce reliance on ad-hoc prompting and establish a repeatable path from system model to deployed code.
By transitioning from prompt-driven to specification-driven development, we provide a methodology to harness the speed of Agentic AI without sacrificing engineering discipline. We address two critical questions for the future of digital engineering: Is it feasible to use MBSE artifacts as the primary source of truth for autonomous coding agents? What specific architectural patterns must a behavior specification follow to ensure deterministic software synthesis?
