Reference
What is agentic CAD?
A working definition, written by people building one. Agentic CAD is software that takes a mechanical design job from a written requirement through to a set of drawings someone can sign, and does the steps in between without being driven click by click.
Last reviewed 10 September 2026 · about a nine minute read
The short definition
Agentic CAD is computer-aided design software in which an AI agent holds the goal rather than the mouse. You state what the part has to do. The system decides what geometry satisfies that, builds it, works out which analysis would prove or disprove it, runs that analysis, reads the result, changes the design if the result is bad, and produces the documents at the end.
The word doing the work is agentic. An agent chooses its next action from the state it is in, rather than executing a fixed sequence. In a design context that means the system decides that a bracket needs a rib, not because a rib was requested, but because the deflection it just computed was too large.
The test is simple. If a person has to look at the result and decide what happens next, it is a tool. If the software decides and can say why, it is an agent.
What it is not
Four adjacent things get called agentic CAD, and none of them is. The distinction is not pedantic: they fail in different places and they are worth different amounts.
| Category | What it does | What it leaves you |
|---|---|---|
| Text‑to‑CAD | Turns a sentence into geometry, usually a mesh or a single feature tree. | A shape nobody has checked. The engineering has not started. |
| CAD copilots | Suggests the next feature, names your sketches, finds a similar part. | You still drive. It saves keystrokes, not decisions. |
| Generative design | Optimises a shape against loads you define, inside a space you define. | An organic solid, a meshing problem, and all the documentation. |
| Simulation automation | Runs a study you set up, on a model you supplied, on a schedule. | Results. Reading them and acting on them is still yours. |
Each is useful. None of them closes the loop, and the loop is where the hours go. A designer's day is not spent extruding. It is spent deciding whether the thing is good enough, and then producing the paperwork that says so.
The four things it has to do
For a system to deserve the name, four capabilities have to be present in one loop. Three of them exist separately today in mature products. It is the joining that is new.
Build real geometry
Not a mesh. A parametric solid with a feature history, exportable as STEP, that a meshing tool will accept and a person can open and edit afterwards. Geometry built to look right and geometry built to be solved are different things.
Choose and run the study
Deciding that a part needs a modal analysis rather than a static one is an engineering judgement, and it comes before any solver runs. The agent has to make it, set the boundary conditions, and mesh to a density it can defend.
Read the answer honestly
The hardest of the four. A stress at a sharp internal corner has no converged value, and reporting one is a lie. A requirement that could not be evaluated is not a pass. A system that cannot say "I do not know" cannot be trusted when it says "it holds".
Issue the package
Dimensioned drawings, a bill of materials, inspection sheets and a report, generated from the same model that was analysed, so the paperwork cannot drift away from the part it describes. This is most of the calendar time on a real job.
Why it is hard
Language models are good at generating plausible things. Engineering is a field where plausible and correct are far apart, and where the cost of the gap is borne by whoever signs the drawing.
A wrong part still builds
Geometry that compiles is not geometry that is right. A fillet placed on the wrong edge produces a valid solid of about the right mass. Nothing raises an error. The only defence is to check a property of the result against a figure derived independently of the code that produced it.
Solvers answer questions they were not asked
A finite element solve with the boundary conditions wrong returns a number, not a warning. Remove the constraints entirely and many solvers report a rigid body displacement of some enormous value and exit successfully. Reading exit codes is not validation.
Refinement does not always converge
The peak stress beside a clamped face or a sharp re-entrant corner rises without limit as the mesh gets finer, because the exact solution there is singular. A system that refines twice, sees the number move, and reports the finest value is reporting the mesh rather than the part.
Units are where silent errors live
Millimetres with kilograms is off by a thousand in force and a million in stress, and nothing complains. Degrees and radians share a dimension, so a check on dimensionality passes between them and the conversion then scales the value by fifty-seven.
None of these produce an error message. They produce a confident wrong answer, which is the worst output an engineering tool can have.
How to evaluate one
If you are assessing a product in this category, these are the questions that separate a demo from a tool. They are the ones we would ask.
- Bring your own part. Ask for a part you already have drawings for and already know the right answer to. A demo on the vendor's chosen geometry tells you what the vendor chose.
- Ask what happens when it cannot do the job. A system with no not-evaluated state has no way to be honest with you, and will pass something it should have refused.
- Ask how the analysis was checked. A hand calculation from a published formula, computed before the solve, is a real cross-check. Agreement with the system's own earlier run is not.
- Ask what the mesh study was. One mesh and a number is not a result. Several densities with the answer visibly settling is.
- Ask what comes out. STEP, DXF and PDF are what every CAM and CAD system reads. A proprietary format means the output is only as durable as the vendor.
- Ask where your geometry goes. For defence, aerospace and anything under an NDA, whether the model leaves the building is a procurement question before it is a technical one.
- Ask to see a figure traced back. Pick a number in the report and ask which run produced it. If nobody can answer in under a minute, the report is prose.
Where the category actually is
Honestly: early. Text-to-CAD works and is shipping. Simulation automation is decades old and mature. Copilots are appearing inside every major CAD package. The part nobody has finished is the loop that connects them and the judgement that has to sit in the middle of it.
Anyone claiming a solved end-to-end system today is describing a roadmap. That includes us: LMCAD is in beta, and it goes on sale on 1 January 2027. What we will do before then is run a part of yours in front of you and let you watch where it struggles.
We think that is the only demo worth anyone's time in this category, because the interesting question is never whether it can make a shape. It is whether you would sign the drawing.
Bring us a part you have already built
One you have drawings for, where you already know the right answer. We run it while you watch. If it gets it wrong, you will see that too.
Book a demo