AI-Enabled Systems Engineering

MBSE and agentic AI need each other in a specific, mechanical way, worth stating plainly rather than treating as a vague synergy. MBSE gave systems engineering a structured, machine-readable model instead of a pile of documents, but building that model has stayed stubbornly manual. Creating a SysML v2 model is still largely hand work, even with a better modeling language underneath it (Rafique et al., 2025). Agentic AI is exactly the kind of tool that's good at fast, structured, repetitive drafting work. Put the two together and each one's weak point is the other one's strength.

The mechanical reason this is possible now, and wasn't really possible with SysML v1, is that SysML v2 has a textual notation alongside its graphical one, and ships with an open API and services specification (Object Management Group, 2025). A model an AI system can parse and generate as text, and reach programmatically through a real API, is a fundamentally different integration problem than a model locked inside one vendor's graphical tool. That's the substrate the rest of this lesson depends on.

The most mature application right now is turning natural-language requirements directly into model elements. Researchers presenting at INCOSE's 2025 International Symposium, in a track specifically dedicated to large language models for systems engineering, demonstrated generating SysML use case diagrams directly from requirement specifications, the first modeling step in a typical MBSE workflow (Schleifer et al., 2025). Separately, engineers at MITRE built AI and natural-language-processing tooling to automate deriving, decomposing, tracing, rewriting, and evolving requirements at the volume real programs actually produce them, work that's manual, time-consuming, and error-prone by hand, with costs that compound the longer an error goes unnoticed (Cannizzaro et al., 2025).

Two things need to happen for a model to be useful: it has to get built, and it has to stay right. Generative AI is being used to accelerate the first part, turning sparse scans or design data into usable 3D models and parameter sets rather than requiring a modeler to build geometry from scratch (Gebreab et al., 2024). On the simulation side, large language model agents are being used to parametrize and adjust simulation models automatically, searching for feasible settings instead of requiring an engineer to hand-script each change (Xia et al., 2024). Both are the same basic move: hand the mechanical, structured part of model construction to an agent, and keep the judgment calls with the engineer.

If there's one place the MBSE-plus-AI partnership is furthest along in practice, it's the digital twin. A 2026 paper frames this directly: an agentic digital twin is what you get when a twin's model-based structure is paired with an AI layer that can reason over the twin's data and act, not just display a dashboard (Ivanov, 2026). That's a three-way relationship, not a two-way one.

Programs are already building toward this combination directly, not just proposing it in a paper. A nuclear-material testbed presented at INCOSE's 2022 symposium, named Beartooth, pairs an MBSE model, a digital twin, and extended reality, the substrate this agentic pattern sits on rather than a finished instance of an agent reading the twin and writing validated changes back into the model (Houck et al., 2022).

Two honest limits are worth stating before this sounds more finished than it is. First, cost. Specializing a general-purpose model for systems engineering work, through fine-tuning or retrieval-augmented generation, is its own engineering project with its own tradeoffs between accuracy, latency, and expense, not a setting you toggle on (Wach et al., 2025). Second, and more fundamental, scale. Generating a single use case diagram from a short requirement plausibly sits inside current agents' proven reliability horizon. A full architecture trade study across a real system's requirements does not. As of early 2025, frontier models completed tasks up to just under an hour of human-equivalent effort about half the time, a number that has grown quickly but still sets a real boundary on how much of a systems engineering workflow can currently be handed to an agent end to end (Kwa et al., 2025).

Put together, the mechanism is straightforward even where the results are still early. SysML v2's textual notation and open API give an agent something concrete to read and write. The agent's perceive, plan, act, observe loop gives it a way to work through a model methodically instead of guessing at unstructured text. And the governance question that loop always raises, how much runs without a person, is exactly as unresolved here as it is anywhere else agentic AI operates.

The remaining lessons in this series, on systems of systems, the digital thread, requirements engineering, and systems architecture, are the specific disciplines where this partnership is currently being tested hardest.