A comprehensive survey published on Hugging Face introduces the concept of agentic artifact creation, defined as the stateful, feedback-driven construction of deliverables by AI systems. The authors reviewed 259 works available as of August 20, 2026, including 230 systems that fit the definition and 29 benchmarks for evaluating such construction.
The survey frames the process as linking three core components: an operational representation of the artifact, a construction policy, and runtime verification whose feedback redirects subsequent actions. This framework is applied across six artifact families: textual, vision, audio, video, spatio, and behavioral.
Key findings highlight that construction challenges depend not only on modality but also on how tightly decisions are coupled and whether failures become visible while still repairable. The authors note that decomposition can reduce local complexity but increases coordination and reassembly costs, and that learned judges may add little independent evidence if they share the generator's biases.
The survey proposes principles for maintaining accountability, including keeping commitments explicit, turning feedback into targeted repairs, and revalidating affected state after changes. It also identifies opportunities for sustaining coherent control as artifacts and systems evolve. A curated list of papers is available on GitHub.