Design workflows for CAD agents: what works, with evidence (verified 2026-10-27)

Scope: how the best agentic CAD systems, professional mechanical-design practice and model-integrity verification can be turned into a process that a CAD agent executes. landscape.md covers tool features; this survey covers workflow and method. The plan that applies it is docs/design-process-plan.md.

Method: papers were read on arXiv (abs and HTML, version and date noted); repositories through the GitHub API and raw files; vendored upstreams (vendor/) from their pinned source. (V) = verified from the primary source; (u) = README, vendor page or third-party claim not checked further. Licences are noted before any idea or code is borrowed (§7). Machine-readable copy: design-workflows.json.

Motivating incident, stated generically: a turret assembly in a robot project rendered its ammunition outside the hopper that should hold it, with subassemblies that read as boxes stuck on the side. Every per-part check passed. §6 says which practices would have caught it.

1. Summary: the practices that have evidence

2. Agentic and AI CAD systems: workflow and method

2.1 Pan-Chera/Multi-Agent-CAD (MAC): the judge agent

• github.com/Pan-Chera/Multi-Agent-CAD, MIT (V, API), last push 2026-10-17, 1,055 stars. Vendors text-to-cad's runtime under MIT.
• Pipeline (V, multi_agent_cad/README.md, WORKFLOW.md, prompts/qa_judge.md): Spec Planner (prompt and reference images → CADBrief) → Geometric Architect (ArchitectPlan) → a deterministic translator from plan steps to build123d (unsupported steps become TODO markers for a repair agent) → execution → two QA engines (STEP topology; STL mesh) → QA Judge → Aider-based repair. Outer loop of 5 retries; the judge engages only after one blind repair; inner loop of 3 for execution failures.
• What the judge sees (V): the user request as ground truth, the planner's special_features, per-feature bounding boxes measured before booleans (white-box instrumentation), the QA engines' error list, the retry count, and, only with JUDGE_MULTIMODAL, four 512×512 isometric renders covering all octants. The prompt says: "You never see the raw 3D mesh, STEP file, or Python source" and "When view evidence conflicts with feature_measurements, trust the measurements."
• Decision schema (V): {action: accept|repair|halt, reason, confidence, evidence[], disputed_errors[], semantic_verification: verified|failed|unverified, modification_suggestions}. Accept or halt requires at least one hard data point in evidence, otherwise the pipeline downgrades to repair; hedged words are not evidence; without images semantic_verification must be unverified, and lack of vision cannot trigger a repair.
• Own evaluation (V, docs/quantified_quality.md): 10 prompts, 141 binary features read off the prompts. A single-agent skill baseline scored 138/141 at 104 M tokens; MAC v1 140/141 at 0.9 M tokens (116× fewer); MAC v2 with visual and geometric judging 140/141 at 1.4 M tokens. The document itself says who judged is unstated, the comparison is "controlled but not isolated", and assembly outputs are curated. The robust finding is cost, not quality; 3 vs 1 failures on 141 items is noise.

2.2 earthtojake/text-to-cad cad skill, build123d-mcp and agentcad (vendored, pinned)

• text-to-cad (repo, MIT (V, vendor/text-to-cad/LICENSE), pinned at 2f79365, imported 2026-10-24; upstream released 0.8.1 on 2026-10-27). Workflow (V, skills/cad/SKILL.md, references/): a model is a parameterless decorated function; the rebuild is the script run; user-given dimensions are checked by an assert in the model or a measurement, "never read off a picture." Snapshot policy (V, snapshot-review.md): one reviewed snapshot per visible change, views chosen for the feature, "No fixed set of views proves every face or feature is correct", "Do not loop on snapshots", every visual concern becomes a deterministic check (asymmetric holes → measure centres; floating boss → solid count and contact distance; wrong bore → clipped view then measure). Inspection recipes (V, inspection-and-validation.md): overlap_volume (zero means at most touching; intended fits must be named), closest_points (does not measure penetration), cheap bounding boxes as a pre-filter, is_valid, self_intersections, boundary_edges; a check that cannot decide raises and "has shown nothing about the design". A scene is bound to a document_hash; selector ids are revision-scoped. No plan-before-build step exists (V): the brief says a simple part may need a sentence and a complex assembly a dimension/interface checklist, with no required format.
• build123d-mcp (repo, Apache-2.0 (V, LICENSE)). Its system prompt (V, default_prompt.md) orders the loop: measure after every execute (faces and volume must change as expected after a boolean), clearance between mates before any render (touching or apart, never interpenetrating), render only after those pass, design_audit (±10 % nudges of every named parameter; brittle/coupling verdicts) before finishing, validate then test-export. Tools (V, llms.md): clearance returns one of apart | touching | containing | interpenetrating with intersection_volume, a_volume_outside_b and, for containing, the worst-direction wall thickness; verify_spec (experimental) checks a declared spec with evidence tiers measured | structural | recognised | unverified, and "conforms = no FAILs and ≥1 requirement actually checked"; lint_drawing must return zero violations before a drawing is done. CADGenBench: a build123d-MCP entry leads the leaderboard at 86.9 % (u, via a 2026-10-25 third-party snapshot; the HF Space did not render).
• agentcad (repo, Apache-2.0 (V, LICENSE)). Validation layers (V, validation.py): BRepCheck, shell closure (free edges), mesh manifold, structure, advisory (sliver faces, loose tolerances). Authors declare expect_solids, expect_shells and floating: true per shown object (V, validation_guidance.py); a disconnected body that is not declared floating is reported with its nearest part and distance. edit_risk.py classifies imported B-reps (face/edge counts, free edges, severe validity errors) before an agent commits to editing them. Shared human/agent review comments live in .agentcad/reviews/comments.json with statuses draft | open | addressed | resolved (V, reviews.py). Its CADGenBench claim (48 % higher score, 2.5× valid outputs with GPT-5 Nano) is a single uncontrolled run by its own README (V, README caveat).

2.3 Commercial systems

• Zoo Zookeeper (press release 2026-02-05, V): researches, proposes constraints and parameters and "generates a clear design plan before producing a CAD model"; has engine-level tools to inspect, snapshot and debug geometry; output is an inspectable feature tree. No verification method or metrics published.
• Autodesk Assistant (Fusion): autodesk.com refused the fetch (403). Trade press (u): assistant GA at AU 2026; an agentic version that can "flag problems, trace knock-on effects" previewed for 2027; Fusion MCP servers released June 2026. No published method.
• Onshape AI Advisor (feature page, V): documentation Q&A only; "does not generate designs, predict errors in real time, or make complex engineering decisions."
• Leo AI (vendor page, V as a claim, u as fact): finds guidelines and formulas, runs calculations, asks clarifying questions, outputs an editable assembly with a part and feature tree. No accuracy data.

2.4 Academic systems with verification or decomposition

2.5 What the evidence says

• Kernel feedback is the most reliable signal. CADSmith's no-vision configuration (an LLM judge reading kernel metrics and validity, no render) gives most of the total gain (median IoU 0.81→0.96); CADCodeVerify's compile loop; Embodied CAD and AssemCAD get 100 % / 89 % executability by gating every op on the kernel; CAD-Coder's zero reward for non-executing code cut invalidity to 1.45 %. CADGenBench's hard validity gate is why tool-using entries dominate. Caveat (Embodied CAD): solver acceptance does not prove intent.
• Assembly-level decomposition (interfaces first) has the largest measured effect. ArtiCAD's "late assembly" ablation and AssemCAD's typed spec (text-to-CAD APR 87.5 % vs 66.7 % for the best baseline; one baseline, Claude-4.8, scores 0.83 % APR at 85 % success, which shows how little a high success rate says about assembly structure). For single parts, planners appear in CADSmith, Idea-to-CAD, MAC and Zookeeper but no paper ablates the planner alone; MAC's 99.3 % vs 97.9 % is within noise.
• Render plus VLM self-critique helps modestly and mainly catches gross errors. CADCodeVerify −18 % distance and +21 % without images; CADSmith's judge moved the mean, not the median; CADFusion's visual preference raised the score but also the invalidity rate; ArtiCAD, CADDesigner and Idea-to-CAD never ablate the reviewer. The survey (2505.08137 §7) concludes LLMs "can generally determine whether the design is satisfactory, but they often lack the necessary information to understand how it should be improved."
• Documented visual-review failure modes. CADCodeVerify's VQA answers are only 64.6–68.2 % correct with 19–27 % "Unclear"; its metrics miss "small gaps between a desk's legs and top". CADSmith passed a quadcopter with arm-hub gaps on the first try (fixed 3 views, render scale). CADReview drops on hidden internal components. BenchCAD: a target render replacing text scores near zero (renders carry no numbers); adding one to text barely helps and hurts the strongest thinking model. CADFusion uses one view because "LVMs perform worse with multiple images." 2607.05573: a VLM critic diverges most on rotationally symmetric parts (u, abs). CAD-Assistant: most failures were reasoning or "misreading tool-generated renderings." Both MAC's judge prompt and text-to-cad's policy encode the same lesson: renders are for relative shape; dimensions, interference and containment come from measurements.

3. Professional mechanical design practice an agent can execute

For each practice: the canonical reference, the artefact, what an agent can produce or verify from a parametric model plus a YAML dossier, and the honest limit. NPR 7123.1D was read from the published PDF. Paywalled standards are cited, not read. Access failures: DTIC (MIL-STD-1472 tables), ANSI previews, ecss.nl (ESA outage), PTC and FreeCAD wiki pages; values from those are (u).

4. Model-integrity verification: algorithms, APIs, licences, performance

4.1 Contact graphs and interference

• Commercial reference behaviour (V, vendor help pages). SolidWorks Interference Detection reports a B-rep interference volume per pair and has Treat coincidence as interference (touching faces reported), Create fasteners folder (thread overlaps segregated) and Include multibody part interferences; the help warns that faceted curved faces can touch falsely. Onshape's tool reports the interfering region's bounding box, not a volume, and excludes standard content (fasteners) by default. Fusion's Design.analyzeInterference() returns interference bodies with volume and can include coincident faces (zero-volume hits). The Fusion 360 Gallery assembly dataset defines a contact as two B-rep faces coincident or within 0.1 mm (V, docs/assembly.md).
• Liaison graphs. Bourjault (1984) and De Fazio & Whitney, Simplified generation of all mechanical assembly sequences, IEEE J. Robotics & Automation 3(6):640–658, 1987 (V bibliographic; the IEEE page did not serve the abstract): nodes are parts, edges are mating contacts; two precedence questions per liaison give the feasible assembly sequences (u for the question count). This is the graph the connectivity gate builds.
• OCCT primitives (OCCT, LGPL-2.1 with the OCCT exception (V); CadQuery and build123d use it through OCP): BRepExtrema_DistShapeShape(a, b) → Value(), witness points and support shapes, InnerSolution() true when one shape is a solid and the other lies wholly or partly inside it, SetMultiThread(true) (V, refman); BRepAlgoAPI_Common plus BRepGProp::VolumeProperties for exact overlap volume; BRepAlgoAPI_Section for touching faces with zero common volume (u, standard usage); BOPAlgo_CheckerSI for self-interference of one compound with SetLevelOfCheck(0–9) (V); BRepClass3d_SolidClassifier.Perform(p, tol).State() → IN/OUT/ON (V); BRepCheck_Analyzer for validity. Distance 0 cannot distinguish touching from overlapping: combine with Common.
• Mesh-level: trimesh CollisionManager (MIT, V collision.py) wraps python-fcl (BSD, V PyPI; FCL 0.7 with DynamicAABBTreeCollisionManager broad phase, continuousCollide giving a time of contact); manifold3d (Apache-2.0, V) for fast robust mesh booleans that error on non-manifold input.
• Measured here (throwaway experiment, 2026-10-27, CadQuery/OCP on a loaded workstation): 100 synthetic solids (boxes, cylinders, spheres): AABB broad phase over 4,950 pairs → 100 candidates in 3 ms; DistShapeShape on the candidates 9.9 ms per pair; exact Common volume 3.4 ms per pair; SolidClassifier 24 points in 7 ms; contact graph and components trivial; 1.35 s total. Real parts with hundreds of faces cost 10–100× more per pair, which keeps a 100-part assembly within minutes cold (an estimate, not measured: DistShapeShape on freeform or heavily filleted faces and Common on shelled parts can take seconds per pair); a cache keyed by content hashes of the two shapes makes a rebuild that changes one part re-query only its pairs. Script: bench/gate_timing.py.

4.2 Containment

No published "components must lie inside their cavity" validator was found (gap); the primitives are all there:

1. Exact volume test on the B-rep: vol(A ∩ cavity) == vol(A) within tolerance ⇒ inside; 0 < vol < vol(A) ⇒ straddling the wall; 0 with positive distance ⇒ outside. build123d-mcp's clearance returns exactly this as containing | interpenetrating | apart | touching with a_volume_outside_b and, for containing, the worst-direction wall thickness (V, llms.md). The cavity must be a solid: register it from the build, or derive it (hull minus part) and say so.
2. Point sampling with BRepClass3d_SolidClassifier on the item's vertices and centre: cheap, misses thin protrusions between samples; DistShapeShape.InnerSolution() is a one-bit hint. AssemCAD verifies ports this way (V).
3. Meshes: trimesh.Trimesh.contains(points) casts rays in a direction and its negative with parity, accepts only when both agree, retries disagreements once at random, raises on non-watertight meshes (V, source).

4.3 Mates and joints

• Vocabularies (V): Onshape mates (fastened, revolute, slider, planar, cylindrical, pin slot, ball, parallel, tangent, width; limits on the moving ones). build123d joints RigidJoint, RevoluteJoint(axis, range), LinearJoint, CylindricalJoint, BallJoint with connect_to(); values outside range raise; no solver (joints are explicit transforms), so consistency means a tree rooted at a fixed part where every connect_to succeeded. CadQuery assembly constraints (Plane, Point, Axis, PointInPlane, PointOnLine, Fixed, FixedPoint, FixedAxis, FixedRotation) are solved by CasADi/IPOPT as squared residuals (V, solver.py); a non-zero residual after solving is an inconsistent or over-constrained mate.
• JoinABLe / Fusion 360 Gallery assembly-joint (Willis et al., CVPR 2022; repo): code MIT, dataset "non-commercial research purposes only" (V, LICENSE.md). A joint is two occurrences, the B-rep entity pair that defines the axis, offset, angle, flip, and joint_type ∈ {Rigid, Revolute, Slider, Cylindrical, PinSlot, Planar, Ball} with limits. Useful as a vocabulary; the data cannot be used commercially.
• URDF/MJCF: URDF must be a tree (V, spec; check_urdf); inertia must be positive definite with the triangle inequality (common practice, u as a single source); cad-agent's robot.check_urdf already does this.
• Consistency rules an agent can run (synthesis; no published spec found): the joint graph is connected and every part reaches a grounded part; loops close within tolerance; mated axes coincide (angle and point-to-line distance under tolerance); pin/hole diameters compatible with positive overlap length; after placement, every joint's two parts are in contact or within the declared clearance, otherwise it is a floating joint.

4.4 Kinematic interference

• Exact swept volumes exist (Sellán, Aigerman, Jacobson, Swept Volumes via Spacetime Numerical Continuation, ACM TOG 2021; code MIT but patent-pending per its README (V); maintained in gpytoolbox swept_volume, MIT). They are mesh methods, seconds per part, and only needed when the envelope geometry itself is wanted (to cut a clearance pocket).
• Sampled poses are the practice. MoveIt's Setup Assistant builds the allowed-collision matrix from 10,000 random configurations and a 2025 paper notes undersampling can wrongly disable pairs (V, arXiv 2512.23140). FCL's continuousCollide does conservative advancement between two poses (V). MuJoCo: sweep-and-prune broad phase, per-body AABB trees, parent–child filtering, meshes collide as convex hulls (V, docs). SolidWorks Collision Detection stops a drag at the first collision; Onshape animates mates and interference stays a static tool (u).
• Recommendation for ~100 parts × N poses: tessellate once into a CollisionManager; per pose, set_transform the moving subtree and query in_collision_internal and min_distance_internal; pick the pose step from the smallest clearance and the largest surface speed between samples (step ≤ d/v misses nothing, the conservative- advancement argument); confirm any mesh hit on the exact B-rep with Common at that pose (facet chord error gives false positives, the SolidWorks caveat). Report the sample density; never claim an exact sweep.

4.5 Minimum wall and feature size

• Ray casting along the inward normal is what text-to-cad's dfam-check (MIT, V SKILL.md: "fact-only, never pass/fail", mesh only) and Casys mcp-dfm (MIT) do; the documented weakness is that thin diagonal walls are undersampled when the ray follows the normal (V, mcp-dfm). libigl's shape_diameter_function (MPL-2.0) averages a cone of rays, which mitigates it.
• Medial-axis / inscribed-sphere methods are exact: Attene et al., ε-shapes, Computers & Graphics 2017 (arXiv 1704.08049, V): finds thin walls and thin gaps; O(n²) worst case per vertex (2,720 s complete vs 45 s thresholded map in the paper); no code release found.
• Slicers check per layer only: PrusaSlicer's "detect thin walls" offsets by half an extrusion width and takes a Voronoi medial axis (V, OrcaSlicer mirror); Arachne replaces it with adaptive widths.
• B-rep probes: build123d-mcp verify_spec has wall_thickness_at(point, direction) (exact ray through the solid) and an approximate global min_wall_mm (V); OCCT negative offset (BRepOffsetAPI_MakeOffsetShape) and comparison is possible but negative offsets fail on some geometry (V, tracker issues).

4.6 Existing open-source assembly validators and CAD lint

Gaps (V, by absence after search): no open tool checks joint-graph consistency; none checks "part inside its cavity" beyond build123d-mcp's pairwise clearance; none ties checks to a design dossier; none evaluates section or exploded renders as judge inputs. That is the space the plan fills.

5. Visual review of renders by a vision model

What the measurements say (details and citations in §2.4–§2.5):

• Works: a few named views chosen for the question (CADCodeVerify's four orthogonal views; CADSmith's iso, high rear and front); specific yes/no questions written from the spec with "unclear" allowed; a judge stronger than the generator; kernel metrics given to the judge first with the validity gate as a hard fail; measurements trusted over images when they conflict (MAC).
• Does not work: counting many features (VLMCountBench: Gemini 2.5 Flash 0.58→0.30 as compositional difficulty rises, V); small gaps and clearances (CADCodeVerify, CADSmith); interior and hidden parts (CADReview); orientation and unusual viewpoints (3DSRBench ICCV'25: GPT-4o 44.2 %, Claude 3.5 Sonnet 48.2 %, human 95.7 %; orientation 21.6 % / 31.4 %; uncommon viewpoints −13.5 %, V); multi-view reasoning (BLINK ECCV'24: GPT-4V 51.3 % vs humans 95.7 %, specialists +34.6 points, V).
• Unmeasured (gap): no paper evaluates section cuts, exploded views or per-section colour coding as judge inputs. They are the obvious mitigations for the interior and counting failures (AADvark colours faces uniquely, V, but reports no numbers), so the plan uses them and measures them (plan §6).

6. The incident, and what would have caught it

Stated generically: a turret assembly in a robot project rendered its ammunition outside the hopper that should hold it; subassemblies read as boxes stuck on the side; every per-part check (valid solid, watertight, within the bed, no plastic-on-plastic overlap) passed.

flowchart LR
    S1["ammunition outside<br/>the hopper"] --> C1["containment gate"]
    S2["subassemblies touching by<br/>coincidence, or floating"] --> C2["connectivity gate<br/>against declared interfaces"]
    S3["boxes where a mechanism<br/>should be"] --> C3["spec card per part<br/>+ an independent reviewer"]
    S4["nobody looked"] --> C4["review packet<br/>named views, yes/no questions"]
    S5["no record of why"] --> C5["decisions on the cards<br/>+ the dossier lint"]
    classDef here fill:#fffdf5,stroke:#173d3c,color:#173d3c
    classDef near fill:#eef1e4,stroke:#6f8a80,color:#173d3c
    classDef out fill:#ffe1cc,stroke:#bd3a12,color:#173d3c
    class S1,S2,S3,S4,S5 near
    class C1,C2,C3,C4,C5 here

7. Licences before borrowing

Standards (NASA SP-2016-6105, NPR 7123.1D) are US government works and may be quoted; ASME, ISO, SAE, AIAG and ECSS documents are cited by number only.