When does verification feedback actually guide targeted artifact repair?

The survey defines agentic artifact creation as "stateful construction in which an AI system materially constructs or revises a deliverable and intermediate observations redirect later work." Its central conditional is that this control structure "can expose dependencies and support targeted revision, but only when observations identify failures at a scope that the available actions can repair." The claim is that the benefit comes less from having a verification loop than from a match between what the loop can see and what the builder can change. The abstract states this as the conclusion of a review of 259 works available through August 20, 2026: 230 systems and 29 benchmarks.

The mechanism is a three-part functional link: an operational representation of the artifact, a construction policy, and runtime verification whose feedback can redirect later actions. The introduction motivates it with deliverables governed by several acceptance criteria, whose "requirements interact" and whose failures "may be difficult to trace or repair." Its example is Paper2Poster, where a scientific poster must preserve source content, fit a constrained page, and communicate the paper clearly. Those requirements "call for different kinds of evidence; a plausible image or scalar quality score cannot establish that the poster is ready to deliver." Read this way, a single score observes at the wrong scope, because it says the poster is not ready without saying which requirement failed or which action would fix it. Direct generation is said to work for bounded tasks whose outputs are easy to inspect and cheap to regenerate, which is exactly the condition that makes a whole-output check sufficient.

This adds a second axis to Can AI verify research outputs as fast as it generates them?. That note treats the problem as a volume gap, with generation cheap and checking scarce. The survey's condition says that even where verification exists, its granularity must line up with the repair actions available, so a verifier can be present and still fail to redirect the work. The same shift toward verification shows up on the human side in What tasks do expert data storytellers trust to LLMs?, where the checking is done by people in interviews rather than built into a construction loop. The survey's framing also bears on Do generated analysis UIs really work better than chat?. Rigidity there can be read as coarse available actions, though the excerpt never makes that connection.

The excerpt does not establish how the condition plays out in practice. It gives no results from the 230 systems, no measure of how often observation and action scopes mismatch, and no finding on which of the six artifact families handle the match best or what evaluation practice looks like across the 29 benchmarks. Those analyses are announced but not shown. The "only when" is asserted in the abstract, not demonstrated in the passages provided. At the strength the evidence allows, it works as a design question to put to any artifact-building agent, namely what scope its verifier reports at and what actions it can take at that scope, and not as a measured result.