Designing in the Age of AI Agents

Author: Denis Avetisyan

A new agentic system is automating 3D CAD model creation, pushing the boundaries of what’s possible in design and assembly.

AADvark constructs detailed 3D assemblies within FreeCAD by iteratively building upon a foundational rectangular prism, progressively adding components and joints to realize complex models.

This paper details AADvark, an agent-based framework capable of generating functional 3D assemblies, demonstrated by the successful creation of a pair of scissors and various static objects using FreeCAD.

Despite recent advances in artificial intelligence, generating complex, functional 3D assemblies with moving parts remains a significant challenge for agentic systems. This paper, ‘Agent-Aided Design for Dynamic CAD Models’, introduces AADvark, a novel system designed to overcome this limitation by directly reasoning about dynamic part interactions and leveraging external constraint solvers. We demonstrate that by augmenting an agent with specialized tools and visual feedback, we can successfully generate functional 3D assemblies, exemplified by a working pair of scissors. Could this approach unlock a new era of automated design and manufacturing for increasingly complex mechanical systems?

Beyond Manual Design: The Rise of Intelligent Agents

Conventional Computer-Aided Design (CAD) systems, while powerful, frequently demand significant time and highly trained personnel. The creation of even moderately complex designs often necessitates hours of meticulous modeling, drafting, and iterative refinement by skilled engineers and designers. This labor-intensive process isn’t simply about time; it also represents a substantial financial investment in specialized software licenses and, crucially, the salaries of those possessing the expertise to effectively utilize them. Moreover, the steep learning curves associated with mastering these tools can create bottlenecks, hindering rapid prototyping and limiting accessibility for those without formal CAD training. This reliance on manual effort and specialized skillsets presents a considerable challenge to industries seeking to accelerate innovation and respond swiftly to evolving design requirements.

Agent-aided design represents a fundamental shift in how products and systems are conceived and developed, moving beyond the constraints of purely manual computer-aided design. These systems utilize intelligent agents – autonomous entities powered by artificial intelligence – to automate repetitive tasks, explore a wider range of design possibilities, and even generate novel solutions. Rather than simply executing designer commands, these agents can proactively suggest improvements, optimize performance based on specified criteria, and adapt designs to changing requirements. This automation not only reduces the time and effort required for design projects but also empowers designers to focus on higher-level creative problem-solving and innovation, ultimately leading to more efficient and effective outcomes.

Agent-aided design systems are fundamentally reshaping innovation by utilizing artificial intelligence to navigate the complexities of design spaces with unprecedented efficiency. Rather than relying on iterative human exploration, these systems employ algorithms to autonomously generate, evaluate, and refine potential designs, effectively accelerating the creative process. This computational power allows for the investigation of a far broader range of possibilities than traditional methods, uncovering novel solutions and optimizing designs for performance, manufacturability, and cost. The result is a significant reduction in design cycles and the potential for breakthroughs in fields ranging from aerospace engineering to architectural design, ultimately fostering a faster pace of technological advancement.

AADvark iteratively designs 3D assemblies-like scissors-by accepting image or text inputs, generating JSON part and joint definitions, compiling them with a constraint solver, and refining the design based on visual feedback from modified FreeCAD renderings and informative error messages.

Agentic Systems: Iterative Design Refinement

Agentic systems employ Artificial Intelligence Agents to autonomously produce and assess design options. These agents operate by generating potential designs based on predefined parameters and constraints, then evaluating those designs against specified criteria – often utilizing simulation, analysis, or rule-based systems. The process isn’t a single iteration; agents are designed to create multiple design candidates, rank them according to performance, and refine their generation strategies based on the evaluation results. This capability allows for the exploration of a design space significantly broader than is typically feasible with manual methods, potentially identifying novel or optimized solutions. The agents leverage algorithms such as generative models, reinforcement learning, or evolutionary algorithms to create and improve designs without direct human intervention in the creation process.

Agentic systems employ iterative refinement through a continuous feedback loop. Following design generation, verification processes – which may include simulations, testing against specifications, or comparisons to existing data – provide quantifiable insights into performance and adherence to constraints. These insights are then fed back to the AI agent, which adjusts design parameters based on the received data. This cycle of generation, verification, and adjustment repeats continuously, allowing the agent to progressively optimize designs and converge on solutions that meet defined criteria. The efficacy of this process is directly related to the fidelity of the verification methods and the agent’s ability to effectively interpret and act upon the resulting data.

Employing a JSON (JavaScript Object Notation) representation facilitates the manipulation of design parameters within agentic systems by structuring design data into a standardized, machine-readable format. This allows AI agents to programmatically access, modify, and iterate on specific design attributes – such as dimensions, materials, or constraints – as key-value pairs. The structured nature of JSON enables automated exploration of alternative configurations by systematically varying parameter values and evaluating the resulting designs. This programmatic control over design parameters is essential for efficient optimization and automated design space exploration, as it removes the need for manual intervention in adjusting and testing design variations.

AADvark dynamically assembled a functional pair of scissors from a base rectangle, iteratively refining the 3D model across multiple views of a revolute joint (0, 20, 40, and 60 degrees).

Validating Designs: Constraint Solving and Error Messaging

The OndselSolver is a core component responsible for validating the physical realism of CAD Models produced by the AI design process. This verification operates by assessing the generated geometry against a predefined set of physical constraints, including material properties, manufacturing limitations, and functional requirements. The solver determines if the proposed design can exist in the real world without violating these constraints; designs failing this check are flagged for revision. Specifically, it evaluates aspects such as minimum feature sizes, wall thicknesses, stress concentrations, and interference between components, ensuring the resulting CAD Model represents a physically plausible and manufacturable design.

The OndselSolver, during its verification of generated CAD Models, produces specific Error Messages designed to inform the AI agent of design violations. These messages are not generic failures, but rather contain details regarding the specific constraints that are not met, such as dimensional conflicts, material incompatibilities, or manufacturing limitations. This targeted feedback allows the AI agent to iteratively refine its designs, adjusting parameters and exploring alternative solutions until a physically feasible and manufacturable CAD Model is achieved. The granularity of these messages is critical for efficient exploration of the design space and rapid convergence on viable outcomes.

Integrating constraint solving with intelligent agents enables efficient exploration of complex design spaces by iteratively refining potential solutions. The constraint solver rigorously assesses proposed designs against predefined physical and functional limitations, while the intelligent agent utilizes the solver’s feedback – specifically, the identification of constraint violations – to intelligently adjust design parameters. This cyclical process, where the agent proposes, the solver verifies, and the agent adapts, significantly reduces the search space compared to purely random or exhaustive methods. The result is a targeted optimization approach, facilitating the discovery of feasible and high-performing designs within a practical timeframe, even for problems with numerous interacting constraints and variables.

Enhancing Spatial Reasoning with Visual Cues

Despite recent advancements, Vision Language Models often encounter difficulties when tasked with spatial reasoning – the ability to understand and interpret relationships between objects in a three-dimensional space. This limitation presents a significant challenge in fields like design evaluation, where accurately assessing the arrangement and interaction of components is crucial. The core issue isn’t necessarily a lack of visual processing capability, but rather the difficulty in translating raw visual data into a meaningful understanding of spatial configurations, such as identifying overlapping parts, gauging distances, or predicting how a design will function physically. Consequently, even highly sophisticated models may struggle with tasks that require a nuanced grasp of geometry and relative positioning, hindering their effectiveness in practical applications demanding precise spatial awareness.

To facilitate improved spatial understanding, designs rendered in FreeCAD are augmented with deterministic identifiers – unique, consistently applied visual markers. These aren’t arbitrary aesthetic choices; instead, each component within a design receives a distinct, predictable visual tag, allowing the agent to move beyond simply seeing shapes to actively recognizing and tracking individual elements. This systematic labeling provides a crucial bridge between visual input and abstract spatial relationships, effectively translating complex geometric arrangements into a language the agent can reliably interpret and reason about. The consistent application of these identifiers ensures that the agent can accurately pinpoint specific components, even across varied viewpoints and rendering conditions, ultimately enabling more robust design evaluation.

The ability to accurately interpret how components relate to one another within a design is crucial for effective evaluation, and the implementation of deterministic identifiers significantly enhances this capacity for artificial intelligence agents. By providing clear visual cues that delineate individual parts and their positions within the 3D model, the agent can move beyond simply ‘seeing’ shapes to actively understanding their spatial relationships. This improved comprehension enables the identification of potential conflicts, such as intersecting geometries or insufficient clearances, which might otherwise go unnoticed. Consequently, the agent isn’t merely processing visual information, but reasoning about the design’s feasibility and functionality, leading to more robust and accurate assessments of its overall quality and manufacturability.

From Simple Objects to Complex Assemblies: Demonstrating System Capabilities

The system demonstrates a notable capacity for generating functional computer-aided design (CAD) models across a spectrum of complexity, successfully moving beyond static object creation to fully articulated mechanisms. Initial validations involved the design of a static toddler bed, establishing a baseline for geometric feasibility and structural integrity. This foundation then supported the development of a dynamic scissors assembly – a significantly more challenging undertaking requiring precise modeling of interconnected parts and functional joints. The successful generation of this assembly confirms the system’s ability to not only create 3D representations, but also to conceive designs that are mechanically sound and capable of performing intended actions, suggesting a powerful new tool for automated design exploration.

The creation of a functional scissors assembly demonstrates a significant advancement in the system’s modeling capabilities. This complex mechanism, reliant on precisely engineered revolute joints for its operation, was successfully designed in a mere four iterations. This rapid prototyping showcases not only the system’s ability to handle intricate geometric relationships, but also to maintain dynamic feasibility – ensuring the resulting design is physically plausible and capable of performing its intended function. The swift realization of this assembly highlights the potential for automated design of complex mechanisms, moving beyond static objects towards fully functional, moving parts with minimal human intervention.

Robust performance across a spectrum of designs is ensured through the integration of strong verifiers, critical components in validating each iterative step of the modeling process. The creation of a functional scissors assembly, a complex mechanical system incorporating revolute joints, exemplifies this capability; each iteration required an average of 745 seconds for completion. This process involved substantial computational resources, consuming 18.2 million input tokens and generating 2.2 million output tokens per cycle. The efficiency demonstrated with this assembly, and other diverse examples, suggests a clear pathway toward fully automated design workflows, minimizing human intervention and accelerating the product development lifecycle by systematically ensuring feasibility and structural integrity.

The pursuit of dynamic CAD models, as demonstrated by AADvark, necessitates a reduction of superfluous complexity. The system’s success in generating functional assemblies – a pair of scissors, for example – highlights the power of focused design. G. H. Hardy observed, “The essence of mathematics is its economy.” This aligns perfectly with the paper’s core concept: achieving functionality not through intricate detailing, but through elegant simplification. Abstractions age, principles don’t; a functional design, free from unnecessary elements, endures. Every complexity needs an alibi, and AADvark provides one – a clear path from intention to realized form.

Where Does This Leave Us?

The demonstration of agentic assembly, however constrained to the particulars of FreeCAD and simplified geometries, reveals less a solution and more the precise location of the remaining difficulties. The success is not in building scissors – that is merely a consequence – but in the formalization of assembly as a solvable problem for an artificial intelligence. Yet, the limitations are stark. The system operates within a tightly defined parameter space, a curated world of perfect fits and predictable interactions. Generalization beyond this remains the core challenge; a system capable of tolerating ambiguity, imperfect data, and the inherent messiness of real-world construction is, demonstrably, not yet here.

Further progress demands a reckoning with the unstated assumptions embedded within current CAD paradigms. The very notion of ‘correct’ assembly relies on a level of precision rarely encountered outside of manufacturing. An intelligent system must not simply enforce constraints, but negotiate them, understanding that functionality often trumps absolute geometric perfection. To pursue ever-more-complex algorithms for constraint satisfaction is, perhaps, to mistake the symptom for the disease.

The true measure of success will not be the creation of increasingly intricate models, but the ability to relinquish control. A truly intelligent system will not simply build what is asked, but determine what should be built, based on incomplete information and uncertain objectives. And if that sounds like an abdication of responsibility, then perhaps it is. But clarity suggests that is precisely the point.

Original article: https://arxiv.org/pdf/2604.15184.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/