Abstract:

In mid-2025, generative AI could draft text and answer questions. In 2026, agentic AI systems autonomously operate browsers, applications, and learning management systems, execute multi-step research and development tasks, and generate validated scientific computing applications, all without human intervention beyond an initial prompt. This rapid transition from passive text generation to autonomous tool-using agents has profound and immediate consequences for STEM education, research, and assessment.

This paper describes the design, implementation, and educational deployment of a personal agentic AI system that links AI interfaces on the web, desktop, and mobile phone to agentic browsers, local applications including MATLAB, and free local large language models, using MCP servers and a lightweight command server exposed through a free ngrok tunnel. MATLAB serves as a computational and orchestration hub, connecting frontier cloud AI models to scientific computing, data analysis, web automation, and macOS application control.

Several educational applications are presented. First, automated course content generation: the agent downloads interactive MATLAB Live Scripts from a library of over 150 open-source physics, astronomy, and engineering submissions on the MathWorks File Exchange, studies each package, and generates Canvas LMS quizzes with content-appropriate questions. Second, rapid application development: the agent converts a MATLAB Live Script into an interactive HTML5 prototype and then a published iOS application, completing App Store submission with minimal human guidance. Third, agentic scientific computing: the agent autonomously builds a radiation transfer model prototype using professional spectroscopy databases, and validates its output against established tools. Fourth, AI-to-AI teaching: a frontier model trains a free local model to operate MATLAB and macOS applications, with the interaction captured as reusable training data, opening a new direction in educational research.

The paper also addresses two urgent disruptions. Agentic browsers can now complete all assignments in a learning management system from a single prompt, rendering traditional anti-plagiarism measures insufficient. Practical countermeasures including proctoring tools and a rethinking of assessment philosophy are discussed. Separately, new ADA digital accessibility compliance rules that took effect in January 2026 have strained institutions relying on legacy remediation tools. Agentic AI that can read, describe, and reformat any content into accessible HTML offers a faster and more capable path to compliance than conventional approaches.

The tools, skills, and workflows described are documented in detail in open-source File Exchange packages and in posts on the MathWorks Generative AI community forum, and are freely available for adoption by other educators.

Keywords:

Agentic AI, MATLAB, Model Context Protocol, STEM education, learning management systems, assessment, accessibility.