Benchmark: CadQuery vs. OpenSCAD for agentic CAD work

TEXT START: We gave six AI agents the same three printable parts to model, three in CadQuery and three in OpenSCAD, then verified every mesh independently.

The Dissection

This is a tool benchmark disguised as a workflow decision memo. Its real subject is not CadQuery versus OpenSCAD; it is whether an AI agent can produce functional CAD without continuous human supervision.

The result is materially more important than the authors’ platform conclusion: both systems generated printable parts, and independent numeric verification—not visual inspection or kernel self-report—was the actual safety mechanism. ModelRift then uses that finding to justify retaining OpenSCAD and surrounding it with automated QA.

The Core Fallacy

The text treats verification as a permanent human or workflow advantage. Under Discontinuity Thesis mechanics, verification is simply the next cognitive task to automate. Once agents can generate geometry, write parsers, formulate assertions, inspect volumes, test interference, and repair failures, “human oversight” collapses into another software layer.

The benchmark therefore understates its own significance. It does not show that CAD work remains human-controlled. It shows that the production pipeline is becoming closed-loop: specification, construction, validation, correction, and export can increasingly be delegated to machines. The tool choice is operational trivia beside the ownership question.

Hidden Assumptions

• Six runs across three synthetic tasks are treated as meaningful evidence about agentic CAD generally.
• A valid mesh is treated as proof of a correct, functional design; physical printing, fatigue, assembly, tolerances, and downstream use were not independently tested.
• The custom skills, parsers, expected volumes, and numeric checks are treated as neutral infrastructure rather than substantial engineering labor supplied by the benchmark authors.
• “Unattended” is overstated. The agents operated inside a heavily engineered harness with task constraints, version caps, tool-specific skills, renderers, and external verification.
• Numeric assertions are assumed to encode the relevant truth. They can verify declared properties while missing an incorrect specification or an omitted requirement.
• The observed failure modes are assumed to remain stable across models, libraries, kernels, and more complex assemblies.
• OpenSCAD’s speed, sandboxing, and text format are treated as durable platform advantages, despite the possibility that competing runtimes and tooling will erase them.
• The article evaluates success by output quality and pipeline convenience, while ignoring who owns the agent, the CAD corpus, the manufacturing capacity, and the resulting productive surplus.
• Moving humans from modeling into QA is implicitly treated as a stable role. Under P1, that role is exposed next.

Social Function

Primary classification: transition management and partial truth, with prestige signaling and elite self-exoneration embedded in it.

The empirical core is useful and unusually honest about silent geometry failures. It is not simple copium. But its framing converts labor substitution into an engineering optimization problem. The article makes autonomous production safer, cheaper, and more deployable while avoiding the social consequence: every improvement to the verification loop removes another justification for human participation.

Its practical function is to prepare the platform for machine-run design while presenting the transition as a choice between runtimes, renderers, and assertion strategies. The worker disappears behind the pipeline.

The Verdict

A credible small benchmark with a larger implication than its authors admit. It establishes that current agents can already generate nontrivial functional CAD and that automated verification is the decisive missing layer.

Under the Discontinuity Thesis, this is not evidence that CAD work survives. It is evidence that CAD is being converted into an automatable production loop. OpenSCAD may be the better local engineering choice for ModelRift, but the strategic direction is terminal for ordinary CAD labor: modeling, checking, and repair are all being compressed into machine-owned infrastructure.