A development environment in which AI agents execute code in a real engineering environment and iteratively verify the results is becoming a reality. Two open-source tools released by MathWorks are noteworthy for clearly separating the roles of AI and engineers to implement a trust-based collaborative workflow.

On the 16th, MathWorks released MATLAB MCP Server and MATLAB Agentic Toolkit, open-source tools based on the Model Context Protocol (MCP).

By utilizing these two tools, AI agents can write MATLAB code and run it directly within an ongoing session, analyzing outputs and errors to iteratively improve the code until the correct result is reached.

It adopts a human-in-the-loop structure where engineers are responsible for result verification and final judgment.

The core of this tool lies in the shift in the AI reasoning method.

MathWorks explained that by having the agent directly execute code in MATLAB, it derives execution results based on deterministic computation and numerical analysis rather than probabilistic reasoning.

Engineers can review output values in the same computational environment, compare them with expected behavior, and iteratively improve the workflow.

The target users are MATLAB users, application AI engineers building agent-based workflows, and platform teams managing AI-enabled engineering environments.

Diego Tamburini, Head of AI Practice at CIMdata, stated, “As organizations adopt agentic AI, the focus of interest is shifting from code generation to trusted execution within existing engineering toolchains,” adding that “actual execution and verification underpin trust in AI-based engineering processes.”

The two tools are provided as open-source packages, △Claude Code (Claude It integrates with various agency tools such as Code, GitHub Copilot, OpenAI Codex, and Gemini CLI.

Developers and organizations can identify, extend, and integrate agent-based workflows using MATLAB in their own environments.

Seth DeLand, Generative AI Product Manager at MathWorks, stated, “By enabling agents to execute MATLAB workflows and iterate on them, teams can move from code generated by LLM to executable and testable results.”