Agentic design under manufacturing constraints

Bridging design intent and manufactured outcome

Mechanical product development is art and science combined [1], [2]. Art in the sense that tacit knowledge, experience and taste influence how successfully new products can be imagined. Science in the sense that manufacturing has a long, empirically driven, tradition of testing, evaluating and iterating on how to bring an idea from concept to reality [3], [4].

Every design decision is inseparable from the realities of manufacturing: material choices, tolerance requirements, ease of assembly, and, perhaps most important of all, marginal cost of production [5]. These factors determine not only whether a product can be made, but whether it is commercially viable. The further along the value chain a project is, the more exponentially costly it becomes to address any design issue that wasn’t uncovered early [6]. Yet, even when all necessary engineering expertise is available during early design and development, inventing hardware products remains one of the most complex undertakings a team can pursue.

The implications of design choices for manufacturability are often invisible until it is too late [7], [8]. In the constant interplay between aesthetics, function, and cost, it is the manufacturing constraints that too often go untested until late in development, when redesigns are costly and deadlines are inflexible [9], [10]. As a result, teams face two undesirable outcomes: accept lower profit margins or delay time-to-market by redoing the design. Neither strengthens competitiveness [1].

Historically, software tools that aimed to help engineers explore design spaces have focused on functional optimization, such as reducing weight while preserving load-bearing strength, improving stiffness, or optimizing flow characteristics using simulation-driven topology modifications. These approaches can yield remarkable geometric and topology optimizations. But, they rest on two fragile assumptions:

1. That the generated geometry is manufacturable at all [11]
2. That it can be manufactured within the cost structure the business case demands [12]

The first assumption occasionally holds. The second rarely does. This mismatch has limited the practical adoption of currently available generative design and topology optimization solutions [13].

For generative design to become practical in mechanical engineering work, it must begin not from performance alone but from manufacturing reality, including cost [14]. A simulation of performance is meaningful only if the resulting part can actually be manufactured within target margins. Otherwise, the process remains academic [15].

Encube was founded with the ambition of closing the gap that currently sits between product intent and manufactured outcome. How might we reimagine concept and prototype design work if designs can be generated under real-world manufacturing constraints from just an idea or a simple sketch?

Today, we’re excited to reveal some of the research and progress we’ve made in our pursuit to create an agentic design engineer that doesn’t just conceptually understand mechanical engineering, but is able to explore and reason about geometry under simulated manufacturing constraints.

From generative CAD to agentic design

Recent years have seen important progress in generative models for computer-aided design (CAD). Early research treated CAD as a language: a sequence of modeling operations like sketches and extrusions that could be learned and reproduced. Wu et al. introduced DeepCAD in 2021, demonstrating that transformer architectures could generate CAD-as-code sequences capable of producing valid solids when executed by a geometry kernel [16]. Alam and Ahmed extended this idea in GenCAD by combining transformers with diffusion priors, allowing generation of parametric CAD models conditioned on 2D images [13]. These works showed that parametric design histories, long the backbone of mechanical design, can be learned and generated.

However, such sequence-based models have inherent limitations. They often support only simple operations and lack mechanisms for validity checking. Alam and Ahmed reported that around 10% of their generated CAD programs failed to produce valid geometry, even when syntactically correct [13]. Subsequent efforts, like the Self-Repairing extensions to GenCAD, introduced post hoc repair diffusion stages to mitigate these errors [17]. Jayaraman et al.’s SolidGen (2022) moved beyond sketches and extrusions by autoregressively building boundary representation (B-rep) elements directly: vertices, edges, and faces, allowing more complex topology [12]. Xu et al.’s BrepGen (2024) then employed a diffusion model over structured B-rep latent spaces, achieving unprecedented geometric fidelity [11].

These developments are impressive, but they share a common constraint: they do not reason about geometry. They generate shapes, but cannot interrogate them, understand spatial relationships, or explore and modify them in context. Furthermore, they do not account for manufacturability during generation. To reach practical engineering utility for generative design, we need a feedback-rich execution environment where designs can be explored, inspected, and revised against geometric and manufacturability signals.

The feedback gap in mechanical design

In software engineering, AI-assisted coding is increasingly successful not just because of the rapid advancement in large language models, but also due to the feedback environment they have access to. Compilers, linters, and test suites form a constant feedback loop: every output can be executed, validated, and corrected [18], [19]. AI coding assistants like Codex and Claude Code rely heavily on this feedback mechanism to identify and rectify errors at compile time and runtime [20], [21].

Mechanical engineering and design lacks this infrastructure. CAD systems can model geometry, and CAM systems can simulate manufacturing, but these tools are disconnected and exceptionally manual in their use. Manufacturability analysis is also rarely integrated; it typically requires manual setup in separate environments such as CAM packages or specialized solvers. There is no equivalent of a compiler error when a generated pocket is too deep to CNC machine, or when a fillet radius is too small to cut for available tools [22].

This disconnect prevents generative design from evolving into agentic design. To make an AI model act like an engineer, to explore geometry, understand spatial relationships, and refine design based on real-world constraints, it must exist in an execution environment that speaks the language of geometry and manufacturing.

Our approach: a spatial reasoning and execution engine

We have developed a distributed simulation engine that provides a dynamic feedback environment for mechanical design similarly to what AI-based coding tools offer.

Our system consists of a Turing-complete programming setup, deeply integrated with an analytical geometry kernel capable of NURBS, B-splines, curvature control, and robust feature operations such as chamfers, fillets, freeform surfaces, as well as generating all types of holes. This environment allows an AI agent to not merely generate geometry, but to reason about it spatially and mechanically, and explore, query, and modify a design iteratively until it decides that it is done. With or without a human in the loop.

In practice, this means that an agent can see and evaluate the geometry it creates: measure dimensions, detect edge curvature, identify self-intersections, and test operations in sequence. When an operation fails, for instance when a fillet cannot be solved due to conflicting surfaces, the kernel’s feedback is parsed by the agent, which then modifies its approach, much as a human designer would.

This allows output to be true CAD-as-code, a human- and machine-readable parametric description of the generated design. Each model is a complete, editable program that can be compiled into a boundary representation (BREP) using industry-grade analytical geometry modeling kernels. The system supports all major native CAD formats, including individual part files as well as assembly files. This ensures compatibility with established design workflows beyond what pure STEP-based approaches offer. Generated designs, therefore, can be imported into tools such as Autodesk Fusion 360, PTC Onshape, Siemens NX, SolidWorks, CATIA, and PTC Creo for further development without additional processing.

Manufacturability as a design constraint

While the system enables general 3D spatial reasoning, we have placed special focus on manufacturability for CNC milling, one of the most pervasive and cost-defining processes in mechanical product development. CNC machining dominates prototyping and low- to mid-volume manufacturing, and its geometric constraints profoundly shape product cost structures [22].

Within this research preview, an agent works inside a parametric and analytical geometry execution environment that exposes CNC-relevant signals during modeling. The following is a non-exhaustive list of capabilities: tool diameter and minimum internal radius coupling, length-to-diameter limits and predicted tool deflection, step-down and step-over constraints, reachable surface normals under 3-axis and 3+2/5-axis kinematics, stock and fixture collisions, setup feasibility, multi-setup penalties, and local surface finish requirements. The agent can query these signals while modeling and receives structured feedback when a constraint is violated.

For example, if a deep cavity with sharp internal corners implies excessive stickout or a corner radius below the available tool size based on a given tool library, the agent will surface a tool accessibility alert and propose a change such as increasing fillet radii, relaxing wall angles, or splitting the feature across additional setups. When predicted material removal time, or operation count, spikes in a region, the agent can report a cost-driver hotspot and investigate alternatives like rest-roughing strategies, different entry conditions, or a modified workholding plan. The modeling loop therefore incorporates physics-aware checks tied to real CNC process variables, rather than relying on abstract optimization targets or statistical priors alone [22].

This work does not claim to solve manufacturability generally. It shows that manufacturability constraints for CNC milling can be brought into the design loop in real time, and that a feedback mechanism can guide the agent to propose geometry that is more consistent with machining realities and cost sensitivities. In practice, we observe that when the agent can reason about both geometry and manufacturability, it tends to converge toward designs with a higher likelihood of satisfying both aesthetic and economic constraints.

Agentic design in action

The core of this work is enabling an Agentic Design Engineer, a system capable of generating, validating, and improving designs through interaction with its environment. The agent iterates autonomously:

• It drafts a parametric design program
• Executes it within a geometry kernel to obtain analytical 3D geometry
• Inspects the geometry, measuring validity, tool reachability, and constraints satisfaction
• Refines and repeats until convergence

During iteration, the agent maintains explanations and deltas: which constraints failed, which geometric edits were attempted, tracks changes such as sharp external edges converted into 45 degree chamfers, and why a particular resolution was selected. Typical edits include increasing local fillet radii to meet cutting tool requirements, relocating or re-sizing holes to avoid clamping regions, or re-orienting faces to enable 3+2 axis accessibility. The result is not a static 3D mesh, but a fully editable parametric CAD model ready for prototype production or continued design using traditional engineering software tooling, such as CAD, CAM or CAE.

In the image above, the agent has modeled a circular mounting plate with a half open pocket that has sharp internal corners that cannot be CNC machined. The execution environment identifies this, contextualizes this insight and feeds it back to the agent to solve the problem. In this case by adding a vertical fillet with a radius that can be machined using the cutting tools the agent has access to. These feedback loops run several times a minute, with support for a human in the loop during more complex design work.

Frontier research: toward hybrid reasoning systems

We view this as a foundation for broader model and system evolution. Future research in Agentic Design should explore combining transformer-based parametric modeling with diffusion-based geometric refinement, merging symbolic design intent with continuous shape representation. However, even with improved model architectures, the feedback principle remains central: an agent should continue to test generated geometry against manufacturability and cost signals to maintain reliability and maximize practical relevance.

Two additional directions appear especially promising:

1. Deeper CNC manufacturability coverage. Extend beyond current analysis and simulation with richer multi-axis kinematics models, fixture synthesis and sequencing, burr and vibrational risk indicators, adaptive tool selection under wear and stability maps, tolerance-driven process planning, and tighter linkage between predicted material removal time and parametric geometry changes [22]
2. A reimagined early-concept user experience. Move beyond the traditional CAD workflow toward conversation-plus-constraint workflows: mixed-initiative sketch and text, live “cost bands” and setup counters that update as parameters change, explainable manufacturability annotations on geometry, and agent-authored design rationales with versioned, human-auditable histories. The goal is to reduce the cognitive distance between intent, geometry, and process during the earliest design work, when changes are cheapest and creativity stand at its highest.

Conclusion

The history of design technology has been one of increasing integration, from manual drafting to parametric modeling, from simulation to optimization. We believe the next major leap is agency: agentic systems that can act, observe, and improve autonomously using both automated and human feedback. This research preview does not claim to close the longstanding gap between design intent and manufacturing reality. It demonstrates that enabling spatial reasoning in 3D, coupling design with CNC-aware constraints, and producing parametric CAD-as-code output that compiles into standard BREP solids, can narrow that gap in a practical way.

If you’re interested in learning more about our research, please reach out to contact@getencube.com, or jobs@getencube.com, if you want to become part of pushing the frontier of mechanical design engineering.

References

[1] C. L. Tan, K. Otto, and K. L. Wood, “Relative impact of early versus late design decisions in systems development,” Design Science, vol. 3, e14, pp. 1–23, 2017, doi: 10.1017/dsj.2017.13.

[2] T. A. W. Jarratt, C. M. Eckert, N. H. M. Caldwell, and P. J. Clarkson, “Engineering change: an overview and perspective on the literature,” Res. Eng. Des., vol. 22, no. 2, pp. 103–124, 2011, doi: 10.1007/s00163-010-0097-y.

[3] V. Krishnan and K. T. Ulrich, “Product development decisions: A review of the literature,” Manage. Sci., vol. 47, no. 1, pp. 1–21, Jan. 2001, doi: 10.1287/mnsc.47.1.1.10668.

[4] G. Boothroyd, P. Dewhurst, and W. A. Knight, Product Design for Manufacture and Assembly, 3rd ed. Boca Raton, FL, USA: CRC Press, 2010.

[5] K. Spütz, G. Jacobs, C. Konrad, and C. Wyrwich, “Integration of production and cost models in model-based product development,” Open J. Social Sciences, vol. 9, pp. 53–64, Dec. 2021, doi: 10.4236/jss.2021.912004.

[6] R. Bakerjian (Ed.) and P. Mitchell (Staff Ed.), Tool and Manufacturing Engineers Handbook, Vol. VI: Design for Manufacturability, 4th ed. Dearborn, MI, USA: Society of Manufacturing Engineers, 1992.

[7] M. Schaffers, E. Rosseel, S. Burggraeve, J. Pelfrene, K. Gadeyne, and F. Petré, “Methodology for early design phase cost and performance trade-off analysis of a new variant in a product family,” Proc. Design Society: DESIGN2022, vol. 2, pp. 593–602, 2022, doi: 10.1017/pds.2022.61.

[8] I. Ullah, D. Tang, and L. Yin, “Engineering product and process design changes: a literature overview,” Procedia CIRP, vol. 56, pp. 25–33, 2016, doi: 10.1016/j.procir.2016.10.010.

[9] T. Xu, J. Xue, Z. Chen, J. Li, and X. Jiao, “A systematic method for automated manufacturability analysis of machining parts,” Int. J. Adv. Manuf. Technol., vol. 122, pp. 391–407, 2022, doi: 10.1007/s00170-022-09586-2.

[10] J. M. Stecklein, J. Dabney, B. Dick, B. Haskins, R. Lovell, and G. Moroney, “Error cost escalation through the project life cycle,” NASA Johnson Space Center, NASA Tech. Rep. 20100036670, 2004.

[11] X. Xu, J. G. Lambourne, P. K. Jayaraman, Z. Wang, K. D. D. Willis, and Y. Furukawa, “BrepGen: A B-Rep generative diffusion model with structured latent geometry,” ACM Trans. Graph., vol. 43, no. 4, pp. 119:1–119:14, 2024, doi: 10.1145/3658129.

[12] P. K. Jayaraman, J. G. Lambourne, N. Desai, K. D. D. Willis, A. Sanghi, and N. J. W. Morris, “SolidGen: An autoregressive model for direct B-Rep synthesis,” Trans. Mach. Learn. Res., 2022. [Online]. Available: https://arxiv.org/abs/2203.13944

[13] M. F. Alam and F. Ahmed, “GenCAD: Image-conditioned computer-aided design generation with transformer-based contrastive representation and diffusion priors,” arXiv preprint arXiv:2409.16294, 2024.

[14] M. Germani, M. Mandolini, and P. Cicconi, “Manufacturing cost estimation during early phases of machine design,” in Proc. Int. Conf. Eng. Design (ICED’11), Copenhagen, Denmark, 2011, vol. 5, pp. 198–209.

[15] A. Siddiqi, G. Bounova, O. L. de Weck, R. Keller, and B. Robinson, “A posteriori design change analysis for complex engineering projects,” J. Mech. Design, vol. 133, no. 10, pp. 101005-1–101005-11, Oct. 2011, doi: 10.1115/1.4004379.

[16] R. Wu, C. Xiao, and C. Zheng, “DeepCAD: A deep generative network for computer-aided design models,” in Proc. IEEE/CVF Int. Conf. Computer Vision (ICCV), Oct. 2021, pp. 6772–6782.

[17] C. Tsuji, E. F. Medina, H. Gupta, and M. F. Alam, “GenCAD-Self-Repairing: Feasibility enhancement for 3D CAD generation,” arXiv preprint arXiv:2505.23287, May 2025.

[18] R. Ravi, D. Bradshaw, S. Ruberto, G. Jahangirova, and V. Terragni, “LLMLOOP: Improving LLM-generated code and tests through automated iterative feedback loops,” in Proc. IEEE Int. Conf. Software Maintenance & Evolution (ICSME), 2025 (Tool Demonstration Track).

[19] G. Dolcetti, V. Arceri, E. Iotti, S. Maffeis, A. Cortesi, and E. Zaffanella, “Helping LLMs improve code generation using feedback from testing and static analysis,” arXiv preprint arXiv:2412.14841, 2024.

[20] Y. Li et al., “Competition-level code generation with AlphaCode,” Science, vol. 378, no. 6624, pp. 1092–1097, Dec. 2022, doi: 10.1126/science.abq1158.

[21] N. Wang, Y. Liu, C. Chen, and H. Lu, “GenX: Mastering code and test generation with execution feedback,” arXiv preprint arXiv:2412.13464, 2024.

[22] A. Pajaziti, O. Tafilaj, A. Gjelaj, and B. Berisha, “Optimization of toolpath planning and CNC machine performance in time-efficient machining,” Machines, vol. 13, no. 1, 2025, doi: 10.3390/machines13010065.