Apply agentic AI to embedded systems

A new line of AI experimentation is moving into embedded engineering, where the question is not whether an agent can generate code, but whether it can operate inside tightly controlled development flows. A May 2026 YouTube feature on agentic AI and model-based design framed the issue around embedded and safety-critical systems, where models, requirements, test artifacts and hardware constraints all have to stay aligned. That is a different setting from consumer coding demos, because the output is expected to survive verification and validation, not just compile. MathWorks describes model-based design as the systematic use of models throughout development, and says the approach is used to develop complex systems with lower risk and greater efficiency. The company’s April 2024 session on “AI with Model-Based Design” said the challenge is how to combine AI with that process for engineered systems. ### Why is embedded work a harder test for agentic AI than ordinary coding? Safety-critical embedded development is governed by lifecycle controls that reach beyond code generation. ISO says functional safety in road vehicles is influenced by requirements specification, design, implementation, integration, verification, validation and configuration, while IEC says IEC 61508 covers the full lifecycle of electrical, electronic and programmable electronic safety systems. RTCA says DO-178C remains the core document for design assurance and product assurance for airborne software, and the FAA says the agency oversees software and airborne electronic hardware used in aircraft systems and in tools used to produce or test installed equipment. (mathworks.com) Those constraints mean an agent is not being asked only to “write firmware.” It is being asked, in practice, to work within a chain of traceable requirements, approved tools, hardware targets and test evidence. That is why verification and validation become central rather than optional. ### What does “model-based” change for an AI agent? Model-based design shifts the unit of work from raw source files to formal or semi-formal representations of system behavior. (iso.org) MathWorks says the method relies on models throughout development, and a 2025 PLOS One paper on safety-critical embedded systems said such projects often require multiple representations, including C, SystemVerilog, timed automata and domain-specific modeling languages. The PLOS paper said manual synchronization across those notations is labor-intensive and error-prone, and presented a framework using bidirectional transformations to keep representations aligned. That points to one of the more plausible uses for agents in embedded flows: not free-form autonomy, but supervised translation, consistency checking and artifact maintenance across representations. That is an inference from the cited material, not a direct claim by the paper. (mathworks.com) ### Why do verification and validation loops matter so much here? TASKING co-CEO Christoph Herzog said in a May 15 EEJournal interview that large language models and agentic AI are being integrated into the company’s toolchain to automate verification and validation tasks. EEJournal said those tasks have long depended on manual processes and that the new work is aimed at safety and security use cases in areas including automotive, aerospace and robotics. (journals.plos.org) That emphasis matches the standards environment. RTCA’s DO-178 training materials include model-based development and verification, formal methods and software tool qualification as part of the certification framework. In other words, the hard part is not only producing an answer; it is producing an answer inside a process that can be checked, repeated and accepted by assessors. ### What kind of AI setup is most likely to work in these systems? (eejournal.com) Constrained tool use is more credible than open-ended autonomy. TASKING said agentic AI can direct external agents within its toolchain, and discussed the Model Context Protocol in the context of maintaining adherence to industry standards. NIST’s AI Risk Management Framework also organizes AI governance around govern, map, measure and manage functions, which fits systems that need explicit controls and monitoring. (rtca.org) In practice, that suggests an agent architecture with narrow permissions, formal inputs, observable tool calls and mandatory review gates. The engineering skill set such systems reward is less about prompting fluency than about decomposition, traceability and instrumentation. That conclusion is an inference supported by the standards and toolchain sources. ### What should engineers watch next? The next evidence will come from vendor toolchains, standards-linked workflows and published case studies rather than from benchmark scores. (eejournal.com) TASKING has already put verification and validation automation at the center of its pitch, while researchers are publishing frameworks for synchronizing models, code and formal representations in safety-critical systems. Future progress is likely to be easiest to verify where companies can show qualified tools, hardware integration steps and test evidence tied to named standards such as ISO 26262, IEC 61508 and DO-178C. (iso.org)