Dark theme
WELCOME TO QQUANTUM

Let Qubi show you around.

Choose how you’d like to explore.

Don’t agreeGo to yourOS instead ↗

Both choices allow Google Analytics, so we can see which pages help.

Sound is your choice. Essential cookies keep the site working. Continuing either way also allows Google Analytics (visits and pages viewed).

Read our Terms, Privacy and Cookies.

Change your choices anytime from the footer.

Getting Qubi ready.
Engineering notes 20 November 2026 By QQuantum.AI 8 minute read
Agentic Ai Computer Graphics Cad Automation Ai Evals Research

Agentic Computer Graphics: What the New Survey Means

A survey from UT Austin, Stanford, SFU, UC Berkeley and Cambridge argues that graphics is code, so LLM agents can write and refine it. It ships no model or code. What the generate, render, check loop means for design and engineering teams.

AI-drafted, machine-checked against our published editorial rules, and reviewed on publication. Corrections: [email protected] How we use AI

Qubi, a blue glass orb with two orbiting electrons, hovers between a cyan code screen and a small turntable in a dark Barcelona studio, where a glowing wireframe vase is being rebuilt layer by layer under a warm lamp.

Article

A new survey, "Advances in Agentic Computer Graphics", argues that because graphics is built on code, the agent loop that now writes and fixes software can also produce 3D scenes, shapes and visual content. It is a map of the field rather than a breakthrough: no model, benchmark or code ships with it. For European design and engineering teams, what matters is the loop's verification step, not the demos.

The paper was posted to arXiv on 1 November 2026 by Liu, Je, Wang and colleagues from UT Austin, Stanford, Simon Fraser University, UC Berkeley and Cambridge, and listed on Hugging Face Papers on 3 November. We researched this piece from the abstract and the Hugging Face listing, not the full taxonomy, so the next two sections stick to what those pages say.

What the survey says

The premise is in the first line of the arXiv abstract: "Computer graphics is built on code." Shaders, scene descriptions, procedural geometry and parametric CAD models are all programs, and the authors observe that "Large language model (LLM) agents can now generate graphics code and iteratively refine it through various feedback."

That sentence describes a specific pattern. An agent writes code, the code is executed or rendered, something inspects the result, and the agent revises. In software the inspector is usually a compiler or a test suite. In graphics it can be a renderer, a geometry check or a vision model looking at the image.

The survey's first stated contribution, again per the abstract, is "a unified taxonomy for characterizing agentic systems across diverse computer graphics tasks". The point of a taxonomy is to let people compare systems built for very different outputs, such as a rendered scene, a mesh or a CAD part, along the same dimensions.

The authors also describe an agent that helps survey the research literature itself. According to the Hugging Face Papers listing, that system is not yet available; the authors say they "plan to release the system publicly". Until it is released, nothing it claims to do can be checked.

The open problems the authors point to

The research summary we worked from records four open challenges raised in the paper: coordinating multiple agents and tools, the interfaces between people and agents, long-horizon planning, and how to evaluate interactive graphics systems. We have not read the full discussion of each, so we treat them as headings, not findings.

They are, however, the same four problems that decide whether an agent system works in any domain. That overlap is the useful part of the paper for anyone outside graphics research.

Our reading: the renderer is the test suite

What follows is our analysis, not the authors'.

Agentic coding works where there is a cheap, trustworthy check. A failing unit test is unambiguous. Graphics is harder because "does this look right" is a judgement, and the obvious automated judge, a vision-language model scoring a render, is itself a model that can be wrong in confident ways. An agent that refines its output against a model critic will get very good at satisfying that critic. Whether it also satisfies a designer or an engineer is a separate question.

So the question to ask of any agentic graphics or CAD tool is what its feedback loop is made of. We would sort the checks into three layers:

  • Deterministic checks. Does the code run, does the model compile, is the mesh closed, do the dimensions match the specification, does the part clear its neighbours. These are cheap, repeatable and auditable, and they should gate everything else.
  • Model-based critique. A vision model comparing the render to the brief. Useful for catching gross errors and for ranking alternatives, unreliable as a final verdict.
  • Human review. A designer or engineer accepting or rejecting the result, with that decision logged so it can become a test case later.

A tool that shows only the second layer in its demos is showing the part least likely to hold up in production. Our guide to evaluating AI systems covers how to build the first and third layers around a model.

Qubi, a blue glass orb with two orbiting electrons, hovers over a workbench in a dark Barcelona design studio, watching a small glowing bracket model pass through a brass measuring gauge while rejected parts lie in a tray under a warm lamp.

Why "graphics as code" suits regulated European teams

The framing in the survey has a practical consequence the abstract does not dwell on. If an agent's output is code rather than pixels, it can be diffed, reviewed, version-controlled and re-run. A parametric model script can be read line by line; a generated image cannot.

For an enterprise team in the EU that is a real advantage. Provenance questions (what produced this geometry, from which inputs, with which model version) have concrete answers when the artefact is a program in a repository. Traceability of that kind is the raw material for the documentation that AI governance and the EU AI Act push teams towards, and it is easier to keep design data inside the organisation when the agent emits scripts that the existing toolchain executes, rather than files from an external generator. Whether a particular obligation applies depends on the use case, and nothing in the abstract or listing we read addresses regulation.

It also changes the cost picture. Each iteration of the loop costs a model call plus a render or a simulation, and a long refinement run multiplies both. Cost per task is driven less by the price of tokens than by how many rounds the loop takes and how expensive each check is. Cheap deterministic checks that reject bad candidates early keep the loop short; leaning on a vision model to judge every round makes it longer and dearer. We set out how to model that in modelling the cost of an AI system.

A worked example, clearly hypothetical

To make this concrete, here is how we would scope a first project. This is our illustration, not a system from the paper and not a client engagement.

Input. A product team wants concept variants of a bracket: mounting holes fixed, overall envelope fixed, material fixed. The brief is written as structured constraints, not prose.

Execution. A single agent writes a parametric CAD script for each variant, in a scripting format the team's CAD tooling already reads. It runs the script, exports the geometry and passes it to the checks.

Verification. Deterministic code checks the hole positions, envelope and wall thickness first. Only candidates that pass are rendered and shown to a vision model, which ranks them against reference images of the house style. An engineer reviews the top few.

Failure handling. If a script fails to run or a candidate breaks a constraint, the error message goes back to the agent with a capped number of retries. When the cap is reached, the task stops and is flagged rather than looping indefinitely. Every rejected candidate is kept, because rejections are the best source of future test cases.

Decision. The team measures how many engineer-approved concepts come out per run, and how long review takes, against its current process. If the approved rate is low, the problem is usually the constraint checks or the brief, not the model.

Note what is absent: multiple cooperating agents. The survey lists agent and tool coordination as an open challenge, and we agree with the implication. Start with one agent and strong checks; add a second agent only when a measured failure calls for it. Our page on multi-agent systems explains when the extra coordination pays for itself.

Where it applies, and where it does not

The pattern fits work whose output is already a program or a structured model: generative concepting for product teams, design-for-manufacture checks, and rule-based BIM quality checks, where much of the verification can be written as deterministic rules. In those settings an agent that drafts and repairs scripts against hard checks is a reasonable thing to pilot.

It fits less well where the only judge is taste, such as brand imagery or marketing visuals. There the loop has no cheap check, and a human remains the bottleneck whatever the agent does. For pure image work, a direct call to an image generation model with a human choosing between results is often simpler and cheaper than an agent loop.

Long-horizon planning, the other challenge named in the paper, is the limit to watch. A loop that refines one part is tractable. A loop that assembles a whole building model or a full product over hours needs checkpoints, resumable state and clear points where a person approves the direction. None of that comes from the model; it is engineering around it.

What we would do now

Nothing urgent. The survey is a useful reading list and a vocabulary for comparing tools, and its literature agent may be worth a look if it is released. For a team doing CAD, 3D or BIM work, the practical step is to write down the deterministic checks its output already has to pass, because those checks are what an agent loop will be built on.

If you want help deciding whether a design or engineering workflow is a good candidate for an agent, and what the checks should be, get in touch and we can scope a short assessment.

KEY FACTS
The paper
"Advances in Agentic Computer Graphics" by Liu, Je, Wang et al. (UT Austin, Stanford, SFU, UC Berkeley, Cambridge) was posted to arXiv on 1 November 2026 and listed on Hugging Face Papers on 3 November.
Core premise
The abstract opens: "Computer graphics is built on code." It says LLM agents can now generate graphics code and iteratively refine it through various feedback.
Main contribution
The authors' first stated contribution is a unified taxonomy for characterising agentic systems across diverse computer graphics tasks.
Literature agent not released
The paper describes an agent that helps survey the research literature. According to the Hugging Face Papers listing, the authors only plan to release the system publicly.
What it is not
It is a survey: it ships no model, product, benchmark result or released code.
Questions

Questions about this project.

What is the Advances in Agentic Computer Graphics paper?

It is a survey posted to arXiv on 1 November 2026 by Liu, Je, Wang and colleagues from UT Austin, Stanford, Simon Fraser University, UC Berkeley and Cambridge. It starts from the premise that computer graphics is built on code, notes that LLM agents can now generate and iteratively refine graphics code, and proposes a unified taxonomy of agentic systems across graphics tasks.

Does the survey release a model or tool?

No. It ships no model, benchmark result or code. It describes an agent that helps survey the research literature, but according to the Hugging Face Papers listing the authors only plan to release that system publicly, so its capabilities cannot yet be checked.

What should a team look for in an agentic CAD or 3D tool?

Look at what its feedback loop is made of. In our view, deterministic checks such as dimensions, clearances and whether the model compiles should gate everything; a vision model critiquing renders is useful for ranking but unreliable as a final verdict; and a human reviewer should approve the result, with decisions logged so they become future test cases.

Why does treating graphics as code matter for European enterprises?

Code can be diffed, reviewed, version-controlled and re-run, so questions about what produced a piece of geometry, from which inputs and with which model version have concrete answers. That traceability supports governance documentation and keeps design data inside existing toolchains. Whether a specific regulatory obligation applies depends on the use case.

Continue reading

Shipped & in build

Real work. Check it yourself.

What we built, what stage it is at, and a link to the live thing wherever there is one.

The updated yourOS homepage: Roxy’s open ring surrounded by bento cards for calls, voice, translation, building, design, apps and more.
The yourOS homepage on a phone, with its compact navigation, welcome film button and bento cards.
youros.tech — the homepage bento hero
Our own · coming soon

yourOS

yourOS is an operating system with an AI at its heart. You talk to it like a person, out loud or typed. It makes your calls, plans your days and works on your projects with a crew of agents, on your own private server. Our own product, built from the first commit; the waitlist is open.

What we built
  • Operating system
  • Roxy
  • Agent crew
  • Voice calls
  • App store
  • Website
Meet yourOS →
The Sonora homepage on desktop: a full-bleed dune landscape behind the headline “Transform Your Life with Sound”, with App Store and Google Play download buttons.
The Sonora homepage on a mobile viewport, showing the same hero reflowed to a single column.
sonora.com — homepage, 1440×900
Live · iOS · Android

Sonora

Website, backend and app assets for Sonora, a free AI sound-therapy app on iOS and Android published by Cognitive AI.

What we built
  • Website
  • Backend
  • App assets
Read the case study →
The updated Podswap homepage on desktop: “Grow on Instagram and TikTok for free, with real engagement from creators your size”, beside a tilted wall of exchange, tool and guide cards on a violet gradient.
The updated Podswap homepage on a phone, with its floating navigation, Instagram and TikTok engagement headline and tilted tool cards.
podswap.app — homepage, 1440×900
Five platforms

Podswap

Brand, website, prototype, apps on five platforms, the AI backend and the algorithm-intelligence pipeline for Podswap, a social growth engine that runs without a human in the loop.

What we built
  • Brand
  • Website
  • Prototype
  • Apps on five platforms
  • AI backend
  • Algorithm pipeline
Read the case study →
The unOpen.AI homepage introducing Roxy, a private AI assistant that remembers who you are.
The unOpen.AI homepage on a mobile viewport.
unopen.ai — homepage, 1440×900
Prototype · closed beta

unOpen.AI

In closed beta: Roxy, a private AI assistant with permanent memory running on infrastructure the user owns. Brand, site, product, backend and conversational design.

What we built
  • Brand
  • Site
  • Product
  • Backend
  • Conversational design
Read the case study →
coherence.com on desktop, headed "Intelligence that moves in phase with people".
coherence.com on a mobile viewport, reflowed to a single column.
coherence.com — homepage, 1440×900
Live

Coherence

Live at coherence.com: brand identity generated with AI agents we directed, the motion system and website for the group behind Sonora, AboutAI, WakeUp and MudShadows. Agent work remains in build.

What we built
  • Brand identity
  • Motion system
  • Live website
  • Agent work — in build
Read the case study →
The new Domain Booth homepage on desktop: the headline “Premium Domains. Serious Buyers. Proven Results.” beside the animated DB monogram.
The new Domain Booth homepage on a phone, reflowed to a single column.
Our build of the Domain Booth homepage — 1440×900
Live

Domain Booth

Live at domainbooth.com for DomainBooth, a premium domain brokerage: a new animated mark, every page redesigned and rebuilt with designed mobile layouts, and a custom CMS for its sales record, featured names, press and enquiries.

What we built
  • Animated mark
  • Every page redesigned
  • Designed mobile layouts
  • Custom CMS
Read the case study →
Who we have built for

Twenty-one years. Names you know.

See all of our work →
  • BBC Good Food logo
  • Charities Aid Foundation logo
  • GAME logo
  • NatWest logo
  • Next logo
  • NSPCC logo
  • Tesco logo
  • Worldpay logo
  • Samsung logo
  • Sony logo
  • Land Rover logo
  • UAP Limited logo
  • Netflix logo
  • Amazon Prime Video logo
  • Mercedes-Benz dealerships logo
  • San Miguel logo
  • Biovene Barcelona logo
  • Seventy7 logo
  • Littlewoods logo
  • Smile logo
  • The Co-operative Bank logo
Let’s talk

Working on
something like this?

If this is the kind of problem you are working on, we are happy to talk it through.

30-minute technical call · no deck