MathWorks extends its MATLAB and Simulink tools with agentic AI frameworks that give AI coding assistants a live connection to MATLAB and specialized domain skills and extend agent capabilities into model-based design. These new frameworks include the MATLAB Agentic Toolkit, Simulink Agentic Toolkit, and MATLAB MCP Server.
AI coding assistants have made it easier for engineers to generate MATLAB code from a prompt, but engineers still have to run the code, interpret the results, and decide what happens next.
Agentic workflows change this division of work. Once an AI model can execute code, inspect the output, respond to errors, and select another tool, it begins participating in the complete engineering workflow.
The expanded role beyond AI coding assistant creates another problem. This access to an engineering environment does not tell the agent which model represents the system, which function fits the task, or if its parameters and validation steps are appropriate. Those decisions depend on engineering knowledge that a general-purpose language model typically does not have.
MathWorks separates these requirements. The Model Context Protocol (MCP) connects the agent to MATLAB and its installed capabilities, and skills provide instructions to the agent, or the MCP server. The question is how much engineering judgment those skills can capture and what should remain with the engineer.

Connecting AI agents to real engineering tools
Connecting a large language model (LLM) to MATLAB gives the agent information that cannot come from its training data alone.
The MATLAB MCP Server gives tools access to identify installed toolboxes, check source code, execute code, and run unit tests. The agent can confirm the capabilities that are available, submit code to the engineering environment, and use diagnostics to guide its next step.
The Simulink Agentic Toolkit widens this access into model-based development, where its tools can analyze model hierarchies, read signal paths, modify model structures, and run behavioral tests. The same modeling environment used by engineers executes the tool calls selected by the LLM.
This extends existing capabilities to AI agents, said Houman Zarrinkoub, principal product manager for wireless communications at MathWorks. "If a human can benefit from MATLAB's algorithms, visualization, and verification environment, imagine agents having the same access," he said.
But access to this toolchain can also expose errors that are hidden by plausible-looking generated code. However, detecting an installed toolbox does not establish that its functions suit the problem.
Static analysis can identify coding issues but it cannot prove that the model represents the target system. Simply passing a unit test only confirms the conditions covered by the specific test criteria.
To address these challenges, MathWorks combines the MCP Server with curated skills in its public MATLAB Agentic Toolkit. These skills can help agents avoid nonexistent functions, identify relevant capabilities, and follow a more direct path through a task, according to the company.
MCP provides access to the engineering environment and returns the results, while skills guide the agent in choosing the appropriate models, functions, and parameters.
Teaching agents how engineers approach a problem
Tool access does not provide a procedure for solving the engineering problem. Skills supply that context.
In the current implementation, each skill is primarily a set of text instructions loaded when the agent identifies a relevant task. Those instructions can identify suitable MATLAB functions, explain parameter choices, and specify the order of operations.
Zarrinkoub explained that some skills guide agents through specific engineering tasks, whereas others could support more complex workflows involving multiple models and analyses.
For example, adding additive white Gaussian noise (AWGN) to a signal requires decisions about the signal-power convention, required signal-to-noise ratio (SNR), and conversions among SNR, Eb/N0, and Es/N0 (normalized SNR metrics). For a specific task like this, a skill can guide the agent through the required parameter choices, conversions, and sequence of operations.
Zarrinkoub also cited a system-level example: a LEO satellite constellation study involving satellite count, orbital altitude, coverage, link performance, and competing objectives. In this application, multiple models and analyses may be required to identify and compare viable configurations. This example shows the complexity that skills can eventually address.
Skills are aimed at reducing the time spent searching documentation, testing unsuitable functions, or reconstructing established procedures. The goal is to help the agent "do it right the first time," Zarrinkoub said.
Customers can also develop their own skills, he said. Internal skills have the ability to capture preferred models, coding rules, and test requirements. This compresses engineering knowledge into reusable guidance.
Digital twins raise the stakes for agent decisions
Digital twins become more imperative when a model is repeatedly reconciled with the physical system it represents.
Zarrinkoub explained this shift in an integrated sensing and communication (ISAC) application, discussed in 6G research. An ISAC wireless network can support data transmission and environmental sensing to help maintain an accurate representation of transmitters, receivers, and surrounding propagation conditions.
Ray tracing can estimate how signals travel through the environment. Increased access to parallel computing hardware has made these calculations more practical, which allows engineers to model environments that were previously too computationally heavy.
The digital twin predicts the system's state using available models and information. The measurements taken from the actual physical environment show where that prediction diverges from observed conditions. Updating the twin model gives the simulation a more accurate starting point that allows engineers to focus on behavior the model failed to predict.
Within this workflow, an AI agent can configure the model, select analyses, run simulations, compare alternatives, and prepare a configuration change for review. Skills provide the domain procedure needed to select the relevant tools and maintain consistent assumptions.

Agent autonomy still depends on human validation
Workflow controls that lower the risk of errors can be combined in an agentic workflow. Permissions and runtime controls determine which files, models, or system settings the agent can change, while skills guide how an agent approaches a task; and tests and simulations determine whether its output satisfies specified conditions.
Instructions inside a skill cannot prevent the model from misinterpreting a procedure or improving one performance target at the expense of another. Those guardrails need to exist outside the skill.
"The final responsibility and accountability is with the human designer," Zarrinkoub said.
Sequential projects carry more risk because each decision establishes assumptions and inputs for later stages, Zarrinkoub said. Early modeling or parameter errors can alter the rest of the workflow.
He proposed dividing sequential projects into defined stages. Engineers first define the design intent and acceptance criteria. The agent completes one stage, after which tests, simulations, and human review determine whether its output can advance. Only an accepted result becomes the input to the next stage.
Review will also require traceability through generated code, tool calls, selected parameters, simulation results, and test evidence. The agent's final explanation cannot substitute for the artifacts needed to reproduce its work.
Zarrinkoub said MathWorks is exploring knowledge bases, knowledge graphs, and specialized agents or subagents that can support complex decisions. He acknowledged that the company is "not there yet."
Engineers still have to define design intent, approve tradeoffs, validate results, and control which actions are permitted. More agent autonomy will only increase the importance of bounded stages, traceability, and enforceable authority.
The post MathWorks Connects AI Agents to Engineering Tools With MCP and Skills appeared first on Embedded.





