2026 ASEE Annual Conference & Exposition

The integration of generative AI into engineering curricula is often constrained by a trade-off between operational efficiency and pedagogical governance: faculty want faster ways to build and update instructional tools, but also need clear oversight of accuracy, structure, and instructional intent. The emerging``vibe coding,'' using natural language to generate executable tools and workflows with AI assistance, can reduce implementation friction and expand what educators can prototype on their own. However, using current AI models as the sole implementation engine for large systems can introduce practical risks around reliability, maintainability, cost, and data governance. As a result, many academic implementations confine AI to post-hoc assistance for discrete coding tasks, limiting educator control over system architecture.

To reduce this efficiency-governance trade-off, this research operationalizes the convergence of three developments: code-centric low-code automation platforms, modern LLMs that generate reliable code when tasks are narrowly scoped, and agentic AI methodologies that decompose complex objectives across specialized assistants.

Building on these developments, we present the ``Professor + AI Team'' model, a faculty-led, Agile-inspired protocol in which the educator acts as product owner and technical lead, coordinating a Development-Layer AI Team for architecture reasoning, targeted implementation, debugging, and example search. We report an auto-ethnographic engineering case study in which a single faculty member, without prior experience in JavaScript, n8n, or vibe coding, successfully architected a 200+ node agentic workflow for curriculum material generation. The case shows that a domain expert can learn and operationalize these tools during development when the workflow is decomposed into inspectable stages, bounded code tasks, and mandatory human-in-the-loop checkpoints.

The protocol emphasizes deterministic scaffolding around probabilistic AI outputs, structured task sequencing, artifact-first state management, and mandatory human checkpoints. Persisted artifacts, including planning documents, manifests, and reusable professor-persona files, function as workflow memory and support cumulative refinement across runs. The protocol also supports selective re-entry and scope adaptation: because intermediate artifacts are persisted as editable files, the educator can modify specific elements, such as adding a learning outcome to the planning artifact or inserting a slide into the slide and notes files, and then rerun only the relevant downstream stages instead of regenerating the entire workflow from beginning.

Using this approach, we observed a substantial reduction in time to a minimum viable product, from an estimated 8 to 13 months for a standard two-person team to approximately six weeks, and we demonstrate replication in a second tool with smaller scope. We also document an emergent outcome we call ``reciprocal competence'': iterative debugging improved the educator's ability to specify constraints, verify outputs, and integrate components. By empowering educators to act as AI orchestrators rather than routine code authors, this work contributes a reproducible blueprint for expert-led educational computing innovation.