What Is Agentic CAD? We Asked CoLab Engineers Building AI for CAD

Agentic CAD has moved quickly from niche research topic to engineering buzzword. It is showing up in vendor announcements, AI demos, product roadmaps, and conversations about the future of mechanical design. It is also one of the fastest-moving categories in the broader landscape of AI tools and agents for mechanical engineers.

Most of that attention is focused on one question: Can AI create CAD geometry from a prompt? For engineering teams, creating geometry is only the starting point. The hard part is what comes next.

Once generated CAD (genCAD) exists, engineers still have to answer many questions that decide whether it can move forward. Can the model be edited in the team’s CAD system? Can it be manufactured at the required volume and cost? Can another engineer understand how it was built?

To date, credible answers to those questions have been few and far between. To help clarify what agentic CAD is and what opportunities exist, we asked two engineers who have built real products and real AI tools for engineers.

Shahed Saleh is a Senior Product Manager on CoLab’s AI team. She is a mechatronics engineer by training, with prior experience building AI-enabled industrial inspection systems at DarwinAI and research roles in additive manufacturing and robotics at the University of Waterloo.

Jeremy Andrews is CoLab’s CTO and Co-Founder. He is a mechanical engineer who ran DFM reviews on Tesla Model 3 battery pack components and at General Dynamics before co-founding CoLab in 2017. He has spent the nine years since building software that reads, renders, and processes native CAD data across 30+ file formats.

We asked Shahed and Jeremy what agentic CAD is, where it works today, where it falls short, and what engineering leaders should expect over the next year and beyond.

What is agentic CAD?

Shahed

Agentic CAD is AI taking goal-directed action inside a geometric design workflow.

The “agentic” part is the difference between advice and execution. A chatbot can answer a question about CAD or suggest what an engineer should click next. An agent can actually drive the CAD tool itself. It builds or edits the geometry, checks the result against requirements, and iterates until it gets there. This workflow is distinct from purpose-built CAD-generation models.

Jeremy

‍I think there are a few different categories getting grouped together right now. One is full text-to-CAD. That is probably the flashiest version, where you prompt a chatbot or a model and it produces CAD data on the other side.

A lot of that work has historically been less about mechanical CAD for product manufacturing and more about 3D models for visual use cases, like animation, video games, or sometimes 3D printing. In those cases, the output may be a mesh or another kind of 3D representation, but not necessarily the kind of parametric CAD data our customers work with.

There are also people working on text-to-CAD for parametric models. That work is interesting, but it still seems early.

The other category is using agents to automate existing CAD tools. The Autodesk and Anthropic Fusion MCP example fits there. That path is probably closer to what actually gets adopted, at least in the medium term, because it works with the CAD tools and file formats companies already use.

What can agentic CAD do today?

Shahed

Useful production use cases today are bounded and lower stakes. This might include CAD documentation copilots, retrieval across file vaults or PLM, repetitive parametric edits in CAD tools, first-pass concept geometry, and review triage that flags issues for an engineer to correct.

The headline version of agentic CAD, where an agent autonomously designs a production assembly, is still mostly demo. It’s quite difficult to get an agent to hold together all the requirements, tolerances, manufacturability, supplier capabilities, and design intent, and keep holding them together across multiple revisions.

That last category is where we started with AutoReview. Checking a design against a team’s own standards and checklists is exactly the kind of bounded, verifiable work agents do well today, because an engineer confirms every finding before it matters.

The most interesting research direction to me is verification. Instead of asking the model to create a perfect design in one pass, newer systems put the agent in a feedback loop with a CAD interpreter. The agent creates a model, gets compiler errors or visual feedback, and iterates. That is promising, but it is not production-grade yet.

Jeremy

Text-to-CAD for real mechanical product development still feels early to me.

There are interesting demos, and there are interesting companies working on the problem as I mentioned. But for actual mechanical CAD models used in product development and manufacturing, it seems further away than the demos sometimes imply.

The more practical near-term category is agents automating existing CAD tools. That could mean using a frontier model to drive the CAD software instead of building a whole new CAD system from scratch. I don’t think that means the problem is solved, but it’s a more realistic adoption path because companies already have CAD tools, kernels, file formats, and downstream processes built around them.

What are the limitations of agentic CAD?

Shahed

Current systems can create something that looks right without reliably knowing whether it is right.

We need to remember that mechanical design depends on context. It’s all about requirements, design intent, tolerances, manufacturability, supplier capabilities, assembly relationships, and prior decisions. Most of that information is proprietary, meaning it lives in company standards, historical reviews, lessons learned, documents, and the heads of experienced engineers. An agent can’t scrape that information from the internet. It has to be captured at the source, which is what CoLab offers during design reviews. Feedback, decisions, and rationale become structured knowledge connected to the geometry it describes.

Without context, a model may generate a plausible part, but engineers know that plausible is not nearly good enough. This is also why I think the geometry gets too much credit. The shape is maybe 20% of the job. The rest is design intent, validation, tradeoffs, collaboration, and the rationale behind the decision. A text-to-CAD demo can make it look like the hard part is turning a prompt into a part. In real product development, the harder question is whether that part should exist in that form at all. The way I see it, engineering is a decision-making process wrapped around geometry.

Jeremy

‍The first limitation comes back to the data question. A mesh model that works for a visual use case is not the parametric, B-rep data engineering teams need to carry a design through simulation, CAM, and manufacturing.

The second limitation is editability. Some new approaches may produce a CAD file, but not one you can load into an existing CAD tool with a full feature tree and keep working on. For engineering teams, that is a major issue.

Will agentic CAD replace CAD designers or mechanical engineers?

Shahed

No. It will change which parts of the workflow deserve time and attention from engineers.

Repetitive CAD work is exactly the kind of work AI can accelerate, such as laying down initial geometry, generating variations, making basic feature changes, or searching for information. The work that requires engineering judgment is not going away.

An engineer still has to decide whether a tolerance is achievable, whether a feature creates an assembly issue, whether a supplier can manufacture the part, or whether the design meets the requirement. AI tools can help with these analyses, but such decisions carry cost, liability, and physical consequences. At the end of the day, an engineer needs to be able to stand behind their work and the decisions that went into building something.

Jeremy

I think the more likely path is AI driving the CAD tools engineers already use.

That means automating more of the design process, but still producing CAD data that engineers can work with in their existing systems.

What is the biggest unsolved problem in agentic CAD?

Shahed

The agent needs access to the information that defines a good design.

For a lot of teams, that information is scattered. At CoLab, we hear this from customers all the time. Design intent lives in requirements documents, company standards, prior design reviews, supplier notes, lessons learned, spreadsheets, PLM records, and people’s heads. We call it the “messy middle” as a shorthand.

A general-purpose model doesn’t know why a previous program made a certain tradeoff, or which supplier is best for a specific part, or which issue caused a lot of rework last time. That specific failure is what our Lessons Learned Agent addresses. It finds similar past designs and surfaces the decisions made on them during a new review, so the work starts from company context instead of internet averages.

Research is also moving toward stronger verification loops, and that direction makes sense. In mechanical design, better feedback signals may matter more than a better agent. Without a reliable way to check correctness, the system cannot be trusted with supporting real engineering decisions.

Jeremy

For me, a big question is whether the approach fits with existing CAD infrastructure.

Most CAD tools use modeling kernels like Parasolid or ACIS to generate B-rep data. Some people think those existing kernels are not the right foundation for AI tooling, and there are companies trying to build more AI-native CAD systems from scratch.

I can see the argument if the goal is to build all-new CAD from the ground up. But if the path forward is AI driving existing CAD tools, then the kernel issue is less central. Those tools are going to keep using the kernels they use, and the AI needs to adapt around them.

So the bigger open question is whether we replace the CAD foundation, or do we use AI to drive the tools companies already have? My current stance is that the second path is more likely in the medium term.

What will agentic CAD look like in 12 months?

Shahed

First, a caveat. This space is moving so fast that anyone who claims certainty is just guessing. Lots of smart people hold different views, and nobody is really wrong until more mature research hits the market.

That said, I expect MCP-style tool-driving to become standard plumbing instead of the headline. The CAD-to-simulation-to-PLM loop will get a bit tighter too. I can see a world where agents get better at moving between tools, executing scoped tasks, and producing first-pass work that engineers can inspect.

I do not expect autonomous design of real products in 12 months. That is probably much further out.

Jeremy

‍I would frame my answer more medium-term than 12 months.

Over the next five to 10 years, I think the most likely path is AI driving existing CAD tools. That could be done by the CAD companies themselves or by third-party models and agents. Either way, it lets teams keep using the tools, file formats, and downstream systems they already have.

What should engineering leaders invest in now to prepare for agentic CAD?

Shahed

We’ve discussed the limitations of agentic CAD, but that doesn’t mean there aren’t some incredible opportunities within reach. As a first step, teams should invest in the knowledge foundation that agents will need. Those investments will pay off even if agentic CAD takes longer to mature than the demos suggest.

Documented standards, structured review feedback, traceable decisions, and accessible lessons learned make human engineering work better today. They are also exactly what context-aware agents will run on tomorrow. That makes them a low-risk, high-upside bet. There is no scenario where a team will regret having its engineering knowledge better organized and connected to its designs.

It’s always exciting to see a great demo, but the value will show up when a team can tell whether the model meets functional requirements, why a decision was made, and who approved it. I think accountability will be more important than just having the most capable agent. If an agent can act inside systems where its work is visible, reviewable, auditable, and grounded in organizational knowledge, then that will be something that truly helps engineers rather than creating geometry that may or may not be useful.

Jeremy

Beyond the knowledge foundation, I would watch two market signals before placing bigger bets.

The first is whether the market moves toward AI driving existing CAD tools or toward all-new AI-native CAD systems.

The second is the difference between explicit and implicit modeling. Most traditional mechanical CAD is explicit. This means an engineer builds the model through a sequence of steps, like creating a sketch, extruding it, cutting material away, and combining features into a part. The model is defined by that feature history, which is part of why another engineer can open the file later and understand or change how it was built.

Implicit modeling works differently. Instead of defining the part through that step-by-step feature tree, the geometry is defined mathematically. That can make it much better for complex forms like lattice structures, where modeling every strut, void, or surface explicitly would be impractical.

I would watch implicit modeling because it may be more naturally aligned with AI-generated geometry, especially in additive manufacturing and highly optimized structures. But the use cases are narrower today. It’s an interesting technology, but it doesn’t automatically solve the mainstream problems around production CAD, editability, downstream workflows, or design review.

Final thoughts

Agentic CAD will increase the pace of CAD creation, but more CAD output does not automatically translate into faster engineering programs. Faster generation will expose the constraints that already slow product development down, such as review capacity, validation workflows, manufacturing feedback, supplier constraints, and late-stage change management. If teams generate more concepts, more variants, and more design changes without improving how those decisions get checked, documented, and reused, the bottleneck simply moves downstream.

The teams that get the most from agentic CAD will be the ones that pair faster generation with stronger engineering judgment. That judgment is informed by earlier checks against standards, easier access to prior decisions, clearer design rationale, and review workflows that improve as more work moves through them.

The advent of agentic CAD is not a harbinger of the obsolescence of engineering skills or judgment. In fact, far from it. It raises the value of experience and judgment that can be found, applied, reviewed, and reused from program to program.

For more on preparing engineering workflows for AI, read building an AI strategy for engineering teams and the best AI design review tools for engineering teams in 2026.

‍