Survey of 259 works maps how agentic AI creation turns one-shot generation into sustained construction

新智元 · wechat · 2026-09-08

A survey from HKUST-GZ, Zhejiang, USTC, Tsinghua, HKU, Sydney and Sun Yat-sen covers 259 works (230 systems, 29 benchmarks) and frames Agentic Artifact Creation as sustained construction: maintaining an evolving deliverable and deciding what to generate, where to edit, and when to stop based on intermediate observations.

Key structure: three functional roles — operational representation (edit granularity), construction policy (next action), runtime verification (render/run/play feedback) — across six artifact classes (text, 2D, audio, video, spatial, behavioral). Core claims: aggregate scores can't guide repair; evaluation must cover final artifact, construction trajectory, and the agent system; multi-agent cost lies in shared state and dependency consistency. Proposes four design principles (encode requirements in artifact state, separate capability from authority, feedback must map to edits, re-verify affected state) and six future directions including global consistency, precise repair, and system self-evolution.

Original post →

More from coding & agent

coding & agent channel →