Agency
More agents do not create more agency.
Agency appears when intermediate evidence can redirect later construction decisions, not when a workflow simply adds roles or stages.
Observation Decision
SurveyOpen catalog2026
Agentic Artifact Creation.
Systems, Evaluation, Principles, and Opportunities
A research atlas of AI systems that create, inspect, and revise artifacts—from documents and images to 3D worlds and software.
- 1 HKUST(GZ)
- 2Zhejiang University
- 3 USTC
- 4Tsinghua University
- 5The University of Hong Kong
- 6The University of Sydney
- 7Sun Yat-sen University
- Papers
- 257
- Named systems
- 218
- Benchmarks
- 28
- Publication sources
- 36
- Dual-axis records
- 220
Interactive framework
A live construction process.
Follow the ordered path from task and policy through representation and verification to the artifact, then back to the next task.
Artifact becomes the next task
An interactive three-dimensional construction process. A task specification enters the construction policy, which acts on an operational representation. Runtime verification observes the result, accepts a delivered artifact, and the artifact can begin the next task cycle.
Research taxonomy
Explore the field.
Start with what an agent creates, or the context where it is used.
Organize by output medium and artifact behavior.
Six familiesOrganize by the real-world context of use.
Six domainsCatalog statistics
At a glance.
Explore how the research is distributed across artifacts, venues, and applications.
Hover or tap to inspect · Use arrow keys when a chart mark is focused
Artifact taxonomy composition
Family totals with their internal artifact-type structure.
Preparing chart...
Publication venue distribution
Rectangle area shows paper count; color groups parent venues by research area.
Preparing chart...
Family growth
Annual counts across the six artifact families.
Preparing chart...
Artifact × application
Coverage density at every cross-axis intersection.
Preparing chart...
Read the chart data · artifact family counts
| Artifact family | Papers |
|---|---|
| Textual Artifacts | 35 |
| 2D Visual Artifacts | 61 |
| Audio Artifacts | 11 |
| Video Artifacts | 34 |
| Spatial Artifacts | 35 |
| Behavioral Artifacts | 74 |
Survey synthesis
What the survey reveals.
Reliable creation connects decisions, state, evidence, and repair.
Decision
Interdependence
Failure
Observability
Diagnosis
Repairability
Feedback
A score is not yet actionable feedback.
Feedback becomes operational when evidence names a criterion, localizes affected state, and reaches a feasible repair or escalation.
Evidence Repair
Validity
Accepted evidence has a version.
When requirements, sources, tools, or artifact state change, dependent approvals should expire before selective revalidation.
Change Revalidation
Proportionality
Control should match construction difficulty.
Stronger explicit control helps when decisions are interdependent, failures are hard to observe, or diagnosis does not map cleanly to repair.
Difficulty Control cost
The open edge Global coherence, targeted repair, and persistent system evolution remain thinly evidenced across artifact families.
Paper library
Search the catalog.
Search by title, author, system, venue, artifact, or application.
257 matching papers
Per page
Sort
-
BenchmarkCodeFlowBench
CodeFlowBench: A Multi-turn, Iterative Benchmark for Complex Code Generation
-
SystemInfiniteWeb
InfiniteWeb: Scalable Web Environment Synthesis for GUI Agent Training
-
BenchmarkSciFlow-Bench
SciFlow-Bench: Evaluating Structure-Aware Scientific Diagram Generation via Inverse Parsing
No matching papers.
Try a broader search or remove a filter to explore more of the catalog.
Reading guide
Understand the field.
What is agentic artifact creation?
Agentic artifact creation uses AI systems to construct an artifact, inspect intermediate results, and revise their decisions using feedback. The survey examines construction policy, operational representation, and runtime verification across different output media.
Which artifacts does the catalog cover?
The catalog organizes outputs into six families: textual, visual, audio, video, spatial, and behavioral artifacts. Examples include documents, images, music, videos, 3D scenes, and executable software. A second axis describes the application context.
How are systems and benchmarks selected?
Entries are maintained in an open, audited dataset. Systems construct artifacts; benchmarks evaluate relevant capabilities. Publication status and taxonomy labels remain visible for each record. See the audit protocol for inclusion rules and evidence.
How can I use or cite this resource?
Start with the survey for the framework, explore the interactive catalog for specific papers, or download the catalog JSON. The BibTeX citation identifies the survey; individual paper links lead to their original sources.
Project
About the survey.
A survey and open catalog of agentic systems for creating, inspecting, and revising artifacts.
Browse papers, inspect the data, or suggest an addition.
Cite this survey
Use the BibTeX entry in your next paper.
@article{wang2026agenticartifactcreation,
title = {Agentic Artifact Creation: Systems, Evaluation, Principles, and Opportunities},
author = {Wang, Tianfu and Hao, Zhezheng and Xia, Xilin and
Liu, Lixin and Hu, Mengkang and Liu, Hongzhang and
Chen, Xi and Liu, Ziyan and Lin, Xiankun and
Zhang, Weijia and Yuan, Nicholas Jing and Xiong, Hui},
year = {2026},
url = {https://arxiv.org/abs/2608.28122}
}