Agent-Aided Design for Dynamic CAD Models

In the past year, researchers have started to create agentic systems that can design real-world CAD-style objects in a training-free setting, a new variety of system that we call Agent-Aided Design. Generally speaking, these systems place an agent in a feedback loop in which it can write code, compile that code to an assembly of CAD model(s), visualize the model, and then iteratively refine its code based on visual and other feedback. Despite rapid progress, a key problem remains: none of these systems can build complex 3D assemblies with moving parts. For example, no existing system can build a piston, a pendulum, or even a pair of scissors. In order for Agent-Aided Design to make a real impact in industrial manufacturing, we need a system that is capable of generating such 3D assemblies. In this paper we present a prototype of AADvark, an agentic system designed for this task. Unlike previous state-of-the-art systems, AADvark captures the dynamic part interactions with one or more degrees-of-freedom. This design decision allows AADvark to reason directly about assemblies with moving parts and can thereby achieve cross-cutting goals, including but not limited to mechanical movements. Unfortunately, current LLMs are imperfect spatial reasoners, a problem that AADvark addresses by incorporating external constraint solver tools with a specialized visual feedback mechanism. We demonstrate that, by modifying the agent's tools (FreeCAD and the assembly solver), we are able to create a strong verification signal which enables our system to build 3D assemblies with movable parts.

Summary

Main Finding

AADvark is a prototype agentic system that can generate dynamic CAD assemblies with movable parts (e.g., a functional pair of scissors) by placing an LLM-based agent in a feedback loop that writes an intermediate JSON assembly, compiles it via a modified constraint solver, and iteratively improves the design using enhanced visual and solver feedback. Key system modifications (quaternions in the solver, deterministic / informative solver errors, and per-instance colored/textured faces and instance labels in FreeCAD) produce strong verification signals that enable the agent to overcome VLM spatial-reasoning weaknesses and produce assemblies with joints and degrees-of-freedom.

Key Points

Problem addressed: prior agentic CAD systems produced only static models and could not specify joints/DoF needed for moving assemblies (e.g., scissors, pistons).
System architecture:
- Agent (e.g., Gemini 3 Flash in demo) ingests images and optional text, writes JSON part definitions and an assembly JSON that lists links (instances) and joints.
- A 3D assembly constraint solver compiles the JSON to a multibody assembly; FreeCAD renders the result. Both outputs (render + solver errors) are fed back to the agent.
Tooling changes that enabled success:
- Modified OndselSolver to use quaternions (avoids Euler flip ambiguity), improve determinism, and provide richer error messages.
- Modified FreeCAD rendering to give unique colors/textures per face/edge and textual instance IDs (Blade1, Blade2, ...), so the agent can unambiguously reference geometry in joint definitions.
- Solver modified to update part positions in FreeCAD even when compilation has errors so the agent can visually inspect failure modes.
Verification focus: designing “strong verifiers” (deterministic solver feedback + enhanced renderings) is central; these verifiers amplify the agent’s iterative correction ability more than changing the agent itself.
Demonstration & results:
- Dynamic case: pair of scissors built from two photos (closed + open). 20 iterations, total 4.14 hours runtime, cost $15.85 (LLM costs reported), 468 LLM calls, 18.2M input tokens and 2.2M output tokens.
- Static cases: toddler bed, chair, end table, carport, whiteboard built from 1–3 images (4–34 iterations depending on object complexity). Toddler bed required the per-face/edge visual encoding to succeed.
Limitations: prototype only supports rectangular-prism parts and a single dynamic joint type (revolute). Agentic execution is nondeterministic and can get stuck; simple restarts often resolve it. Scaling to full FreeCAD feature set and other joint types remains future work.

Data & Methods

Inputs: 1–3 images per object and optional textual specifications or LLM-generated design documents.
Intermediate representation: JSON part files (parametric part definitions) + JSON assembly file listing links (instances) and joints. Joints specify link names, face identifiers, joint type, DoF (fixed/free), and limits.
Agent: multimodal LLM acting in an iterative loop to produce and update JSON files based on renderings and solver messages (demo used Gemini 3 Flash).
Tools:
- FreeCAD (modified): per-instance face/edge unique color/texture + instance textual labels in the renderings.
- Ondsel-based constraint solver (modified): use of quaternions, deterministic numeric routines, richer error diagnostics, and partial updates on compile failures.
Evaluation scenarios:
- Dynamic: pair of scissors from two photos (closed and open). Tracked iterations, LLM calls, tokens, wall time, and cost.
- Static: five household/structural objects (toddler bed, chair, end table, carport, whiteboard) tested for successful assembly generation.
Performance metrics reported: iteration counts (4–34), per-iteration average time ~745 s, per-iteration cost ~ $0.79 (demo averages), overall token and LLM call counts for the scissors run.

Implications for AI Economics

Productivity and labor substitution:
- AADvark demonstrates an agentic, training-free pathway to automate parts of CAD design, especially early-stage prototyping and layout of assemblies with moving parts. This could reduce engineering hours for routine parametric assembly tasks and lower the cost/time of iterative design cycles.
- Current throughput (hours per object, nontrivial LLM token costs) suggests near-term augmentation rather than outright replacement of skilled CAD engineers—useful for rapid prototyping or pre-design drafts.
Cost structure and scaling:
- The demo quantified LLM-driven costs (e.g., ~$16 for a 4h scissors run). As LLM inference costs fall and models become more efficient, per-design costs will decline, making deployment across manufacturing pipelines more economically viable.
- However, high token consumption and many LLM calls per design imply variable marginal costs; firms will need to optimize agent prompts, verifiers, and tooling to make per-item costs predictable.
Value capture and productization:
- High-value returns appear to come from modifying verification and CAD/tooling (solver improvements, richer render feedback). This suggests business opportunities for tool vendors that offer agent-optimized solvers, deterministic APIs, or “agent-ready” CAD renderers—value might accrue to companies that provide these verifiers rather than raw LLM providers.
- Open-source solver/tool modifications (as in the paper) lower entry barriers and could democratize agentic CAD, but proprietary integrations with domain-specific verifiers could become differentiators.
Adoption barriers and quality assurance:
- Limitations (restricted part primitives, only revolute joints, nondeterministic agent behavior) and the need for robust verification highlight risks in industrial adoption—manufacturers will require reliability, certifications, and integration with simulation/FEA workflows before using agent-generated designs at scale.
- The use of strong verifiers (deterministic physics/constraint solvers and informative errors) aligns with an economic pattern where investments in verification infrastructure yield disproportionate improvements in agent performance—firms should budget for solver/tool adaptation when deploying agentic design systems.
Labor-market dynamics:
- The system is likely to shift the demand mix toward higher-skill tasks (design verification, complex constraint definition, system integration) and reduce demand for repetitive CAD modeling tasks. Training and workflows will need to adapt (e.g., engineers supervising agentic loops, curating verifier rules).
Policy and regulatory considerations:
- As agentic CAD affects safety-critical designs, regulators and firms will emphasize traceability, verifiable constraints, and reproducible solver behavior. Deterministic solvers and richer error logs (the paper’s modifications) are directly economically valuable for compliance and liability management.

Short-term takeaway: AADvark shows that engineering effort focused on “agent-friendly” toolchains (verifiers + deterministic solvers + unambiguous visual encodings) produces outsized gains in agentic CAD performance. For firms, the economically sensible approach is to invest in verifier/tool adaptation and targeted agent workflows to capture productivity benefits while managing reliability and cost.

Assessment

Paper Typedescriptive Evidence Strengthlow — The paper presents a system prototype and demonstrations rather than causal or comparative evidence: claims are supported by engineered examples and verification signals but lack systematic benchmarks, statistical evaluation, user studies, or real-world fabrication tests that would substantiate robustness and performance. Methods Rigormedium — Engineering design appears thoughtful — combining LLM code synthesis, a CAD environment (FreeCAD), an assembly/constraint solver, and visual feedback to produce a verification loop — but the description lacks rigorous experimental design, detailed evaluation metrics, ablations, scalability analysis, and reproducibility details that would justify a 'high' rating. SampleA prototype agentic CAD system (AADvark) that operates in a training-free setting by having an LLM emit CAD code, compiling to FreeCAD models, using an external assembly/constraint solver to represent joints and DOFs, and leveraging visual feedback for iterative refinement; demonstrated on example 3D assemblies with moving parts (e.g., pistons, pendulums, scissors-like assemblies) rather than on large benchmark datasets or real manufactured parts. Themesinnovation productivity GeneralizabilityDemonstrations likely restricted to a small set of synthetic or laboratory CAD examples rather than diverse, industry-grade assemblies, Depends on specific toolchain (FreeCAD + chosen solver) and LLM capabilities; results may not transfer to other CAD platforms or solvers without engineering effort, Does not appear to account for real-world manufacturing constraints (materials, tolerances, tolerancing standards, tolerances, machining/assembly processes), limiting industrial applicability, Scalability to large assemblies, complex kinematics, or integrated electrical/thermal systems is untested, Relies on current LLM spatial reasoning and tool reliability; future model differences could change performance

Claims (9)

Claim	Direction	Confidence	Outcome	Details
In the past year, researchers have started to create agentic systems that can design real-world CAD-style objects in a training-free setting, a new variety of system that we call Agent-Aided Design. Innovation Output	positive	high	emergence of agentic CAD systems (training-free)	0.09
Generally speaking, these systems place an agent in a feedback loop in which it can write code, compile that code to an assembly of CAD model(s), visualize the model, and then iteratively refine its code based on visual and other feedback. Other	neutral	high	system architecture / iterative design loop (agent writes code, compiles, visualizes, refines)	0.09
Despite rapid progress, a key problem remains: none of these systems can build complex 3D assemblies with moving parts. For example, no existing system can build a piston, a pendulum, or even a pair of scissors. Other	negative	high	capability to generate complex 3D assemblies with moving parts	0.18
In order for Agent-Aided Design to make a real impact in industrial manufacturing, we need a system that is capable of generating such 3D assemblies. Innovation Output	positive	high	industrial applicability / impact contingent on assembly-generation capability	0.03
In this paper we present a prototype of AADvark, an agentic system designed for this task. Other	positive	high	existence of a prototype agentic system (AADvark) for assembling movable 3D parts	0.18
Unlike previous state-of-the-art systems, AADvark captures the dynamic part interactions with one or more degrees-of-freedom. Other	positive	high	modeling of dynamic part interactions (degrees-of-freedom captured)	0.18
This design decision allows AADvark to reason directly about assemblies with moving parts and can thereby achieve cross-cutting goals, including but not limited to mechanical movements. Other	positive	medium	reasoning about assemblies with moving parts / achieving mechanical movement goals	0.11
Current LLMs are imperfect spatial reasoners, a problem that AADvark addresses by incorporating external constraint solver tools with a specialized visual feedback mechanism. Other	positive	high	spatial reasoning capability improved via external solvers and visual feedback	0.18
We demonstrate that, by modifying the agent's tools (FreeCAD and the assembly solver), we are able to create a strong verification signal which enables our system to build 3D assemblies with movable parts. Other	positive	medium	ability to build 3D assemblies with movable parts (enabled by enhanced verification signal)	0.11