MATLAB’s Unexpected Collision With Agentic AI

When the 42-year-old company MathWorks announced a new feature on X a few days ago, many users reacted with disbelief at the idea of traditional engineering software like MATLAB being integrated with modern AI coding agents such as Claude Code and OpenAI Codex.

The surprise stemmed from the fact that these platforms belong to very different generations of computing. 

For decades, MATLAB and Simulink have been foundational software platforms across academia and industry for numerical computing, simulation, signal processing, control systems, and model-based engineering. 

They are deeply embedded in the workflows of companies such as NASA, Tesla, Ford Motor Company, Airbus, and Siemens, enabling the design of everything from aircraft control systems and automotive electronics to industrial automation and embedded systems.

MathWorks has gradually introduced AI-assisted capabilities over the years, including predictive analytics tools, deep learning frameworks for training and deploying neural networks, and machine learning toolkits for classification, regression, clustering, and data modelling. Some AI features have existed on the platform since the 1990s. 

Its ecosystem already included products such as the Deep Learning Toolbox and the Statistics and Machine Learning Toolbox, which enabled engineers to build AI models long before the generative AI boom.

But irrespective of tradition, legacy engineering platforms are increasingly being drawn into the generative AI era, and MathWorks’s latest step hints at a changing future for engineers and companies deeply tied to MATLAB workflows for years, if not decades. 

Why Bring AI Agents into MATLAB

“With the new Agentic Toolkit, the goal was to connect MATLAB and Simulink with agentic AI tools that are being adopted by a lot of commercial enterprises,” says Seth DeLand, Product Manager for Generative AI at MathWorks, in an interaction with AIM. 

With this toolkit, the company has announced an MCP server that enables an AI agent such as Claude Code, GitHub Copilot, or OpenAI Codex to interact directly with MATLAB. 

Coding agents can run MATLAB code, inspect outputs, debug failures, execute tests, analyse code quality, and iteratively refine programs based on real execution results.

The company says one motivation for the toolkit is reducing hallucinations in LLM-generated outputs. Before the MCP integration, developers often had to manually copy AI-generated MATLAB code into the software only to discover invalid functions, incorrect syntax, or broken dependencies.

Additionally, MathWorks has released a library of ‘Agent Skills’ to teach these systems to perform specific workflows, modernise legacy code, and adhere to set development standards across MATLAB and Simulink.

DeLand says that while MATLAB has text-based code, it was easier to work with LLMs and build integrations, but the company’s core focus was also on integrating agents into Simulink models. 

For instance, Simulink is a graphical block diagram environment for multidomain simulation, and given how it is a crucial part of physical AI systems by providing the virtual physics models required to validate control algorithms before hardware production, it can play a vital role in the physical AI era or the development of hardware that enables AI.

AI Solves One of MATLAB’s Biggest Problems

Beyond code generation and workflow optimisation, the new AI integrations resolve a persistent operational challenge in MATLAB. 

DeLand says that the software's architecture presents an obstacle to discoverability for researchers as MATLAB houses thousands of functions distributed across multiple toolboxes. 

“One of the challenges for us has always been helping users discover new capabilities that we've added in MATLAB,” says DeLand. The integration of AI features addresses this limitation by suggesting code and tools based on problem descriptions. Users previously struggled to find tools through search engines because they lacked the exact terminology. 

The models process the problem parameters and establish connections to the required functions. DeLand explains the mechanism, noting, “it can suggest code, it can suggest functions, it can suggest tools that I probably wasn't aware of”.

Besides, agent integration addresses limitations for developers exploring domains outside their areas of expertise. Historically, professionals required consultations with teams of specialists to understand technologies of adjacency. The new AI features in MATLAB provide immediate pathways into unfamiliar disciplines of unfamiliarity.

“If I am not familiar with signal processing or optimisation or solving partial differential equations, I can get into that space very quickly and get recommendations for what tools to do that in MATLAB,” DeLand says.

He notes that, given how academic projects rely on combining existing technologies to develop novel ideas or research, toolkits facilitate this intersection of systems and accelerate the production of research.

“We’re going to see more and more generative AI features integrated to work with MATLAB,” says DeLand on the roadmap for the company. 

However, he says that existing libraries will remain deterministic, ensuring that companies and legacy institutions can continue to work with MATLAB as before. 

“I can pick up MATLAB code from decades ago, and it still runs—but on top of that, we now have a lot more ways to get AI assistance in creating that code,” says DeLand. 

MATLAB’s Growing Identity Crisis in the AI Era

The MATLAB Agentic Toolkit suggests that MathWorks is beginning to adopt development patterns historically associated with the open-source AI ecosystem rather than traditional engineering software.

The toolkit is distributed through GitHub, updated rapidly, built around the MCP infrastructure, and designed for continuous iteration with AI coding agents.

Notably, these are precisely the areas where MathWorks and MATLAB have long faced criticism from the broader developer community.

For years, developers across Hacker News, Reddit, and Stack Overflow have questioned whether proprietary numerical computing environments still make sense in an era dominated by Python, open-source tooling, and AI-native development workflows. 

Many argue that ecosystems built around NumPy, SciPy, PyTorch, TensorFlow, and Jupyter now replicate much of MATLAB’s functionality without the licensing costs or closed ecosystems associated with MathWorks.

Discussions around MATLAB often centre on interoperability limitations, reliance on proprietary toolboxes, and concerns that MATLAB-centric skills are less transferable to mainstream software engineering than those in Python, Rust, or C++.

The criticism is particularly strong among younger developers, many of whom increasingly see MATLAB as a legacy engineering platform rather than a modern software environment. 

Across forums, developers routinely question whether new projects should continue to begin in MATLAB at all unless they specifically depend on Simulink, Embedded Coder, or deeply entrenched enterprise workflows. Open-source alternatives such as GNU Octave, Scilab, and OpenModelica are also frequently discussed as replacements for portions of the MATLAB ecosystem.

At the same time, MATLAB’s long-term users argue that these comparisons often underestimate what keeps the platform deeply embedded across aerospace, automotive, industrial automation, and research environments. 

Engineers point to Simulink’s dominance in model-based systems engineering, mature code-generation pipelines, specialised domain toolboxes, deterministic behaviour, documentation quality, and decades of validated industrial tooling, all of which are difficult to replace with fragmented open-source alternatives.