# Agentic computer graphics: a survey maps how AI agents draw > A 27-page survey from UT Austin, Stanford and others sorts hundreds of papers on LLM agents that write and fix graphics code into one taxonomy. URL: https://claudegpt.si/news/agentic-computer-graphics-survey Published: 2026-11-04T18:52:10.000Z Author: ClaudeGPT Si Newsroom Section: Research Topics: Research papers, AI agents, arXiv, Hugging Face, 3D generation Publisher: ClaudeGPT Si ## Key takeaways - A survey posted to arXiv on November 1, 2026 defines agentic graphics as language models that act in a graphics environment and use the results to decide their next step. - It sorts systems by four choices: specification, base model, harness and graphics interface, and covers 2D, CAD and 3D scenes. - Among 304 works using observation feedback, 77.0% use visual observations, and local patching rose from 50.0% to 83.0% of systems. Agentic computer graphics is the idea that an AI model can build images, 3D objects and scenes by writing graphics code, running it, looking at the result and fixing it. A survey posted to [arXiv on November 1](https://arxiv.org/abs/2611.01860) by researchers at UT Austin, Stanford, Simon Fraser University, UC Berkeley and the University of Cambridge maps how the field works. It has 128 upvotes on Hugging Face daily papers. The survey matters because it gives a field that grew fast a shared vocabulary. If you build or evaluate tools that make slides, charts, CAD parts or 3D scenes with language models, it shows which design choices exist and which ones people actually use. ## What counts as agentic graphics Graphics runs on code: vector files, Blender scripts, CAD programs, shader code. A language model can write that code. The authors say a system is agentic only if it meets three conditions: - A language model decides at least some of the actions. - The system carries out those actions in a graphics environment. - The results of those actions feed back into what the model does next. One-shot generation does not qualify. Neither does feedback used only during training, or a fixed pipeline where results never change the model's next step. The survey covers 2D media, static 3D content and dynamic 3D content. ## The taxonomy: four design choices The paper sorts every system along four dimensions: 1. **Specification.** What the task is (2D, static 3D, dynamic 3D), what goes in (text, image, video, a 3D representation, a document) and what comes out (a program, a structured spec, an artifact, rendered media). 2. **Base model.** A language model, a vision-language model or a mix, and whether it is used off the shelf, fine-tuned, or post-trained with reinforcement learning. 3. **Harness.** The loop around the model. It covers planning and how work is ordered, how fixes are applied (regenerate, patch locally, roll back, replan, pick among candidates), who checks the result (the model itself, a critic model, a program, the environment or a human) and what memory the agent keeps. 4. **Graphics interface and tooling.** Whether the agent writes code, fills a structured spec, calls tools or clicks a GUI, and what resources it can use: assets, example runs, documentation. ## What the survey found The authors built an agentic literature system to annotate papers at scale. They say it separates what a paper reports from their own interpretation. It is built on ReAct, SWE-agent and PaperQA2, and they plan to release it publicly. The corpus size is stated two ways: "more than 300 papers" through September 2026 in the introduction, and 350 papers from 2024 to September 2026 in a later section. Some of the numbers they report: - Of 304 works that use observation feedback, 77.0% use visual observations, and 58.2% combine several kinds. - Local patching, fixing part of a result instead of redoing it, rose from 50.0% to 83.0% of systems over the period analyzed. - Documentation and rules as a resource reached 85.3% of records in the latest period. - Scene generation leans on visual and structural feedback and asset libraries. CAD leans on runtime feedback, full regeneration and documentation. The applications chapters list systems for SVG and logo animation, design tools, scientific figures, charts, posters and slides, 3D geometry, textures and materials, CAD generation, reconstruction and editing, and full scenes. ## Open problems and limits The abstract names four open challenges: coordinating agents and tools across tasks, designing interfaces for people and agents to work together, keeping long plans on track, and reliably evaluating interactive graphics systems. The survey lists many artifact benchmarks, such as BlenderGym and 3DCodeBench, and interaction benchmarks, such as OSWorld. Treat the percentages with care. They describe the authors' corpus and their annotation, and the paper says its automated checks confirm that required fields are filled and the taxonomy is followed, not that a classification is correct. This article rests on the paper's abstract and the first part of its full text; the later sections on dynamic assets, open problems in detail and the conclusion were not read for this piece. ## What to watch The authors plan to release the survey system. Whether it comes with the annotated corpus will decide how easily others can check these trends. The paper is [on arXiv](https://arxiv.org/abs/2611.01860) under a CC BY 4.0 license, with an [HTML version](https://arxiv.org/html/2611.01860v1). ## Frequently asked questions ### What is agentic computer graphics? It is graphics work done by a system in which a language model chooses actions, runs them in a graphics environment such as Blender or a CAD tool, and uses the results to decide what to do next. ### Who wrote Advances in Agentic Computer Graphics? Hengyu Liu, Jihyeon Je and Ruiqi Wang share first authorship, with co-authors from UT Austin, Stanford, Simon Fraser University, UC Berkeley and the University of Cambridge. ### Is the survey system released? Not yet. The authors say they plan to release the agentic literature survey system publicly. ## Sources - [Advances in Agentic Computer Graphics (abstract)](https://arxiv.org/abs/2611.01860), arXiv - [Advances in Agentic Computer Graphics (HTML full text)](https://arxiv.org/html/2611.01860v1), arXiv