PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic Artifact Creation: Systems, Evaluation, Principles, and Opportunities

Tianfu Wang, Zhezheng Hao, Xilin Xia, Lixin Liu, Mengkang Hu, Hongzhang Liu, Xi Chen, Ziyan Liu, Xiankun Lin, Weijia Zhang, Nicholas Jing Yuan, Hui Xiong

arXiv:2608.28122v1cs.MM

TL;DR

Agentic artifact creation must turn generated components into dependable deliverables despite coupled decisions, late failures, and limited repairability. This survey defines and reviews the paradigm across six artifact families, finding recurring cross-family challenges and principles for explicit commitments, targeted repair, and state revalidation.

  • Problem

    Existing artifact-creation systems face coupled decisions, late or poorly localized failures, and edits too coarse for dependable local repair.

    Source: §1.1 · p. 1, §8.2 · p. 30

  • Method

    The survey defines agentic creation as stateful construction linking artifact representation, construction policy, and runtime verification, then synthesizes evidence across artifact families.

    Source: §3.1 · p. 5, §9 · p. 31

  • Results

    Across six artifact families, challenges extend beyond modality to decision coupling, failure visibility, repair granularity, and preservation of dependencies during revision.

    Source: Summary · pp. 14–16, §9 · p. 31

  • Takeaways & Limitations

    Reliable construction requires explicit commitments and control boundaries, feedback linked to feasible repairs, and revalidation of affected state after change.

    Source: §9 · p. 31

    Cross-family evidence remains too sparse to establish a shared mechanism for invalidation, regression, rollback, and renewed review.

    Source: §7.4 · p. 29

Abstract

from arXiv · show

Generative models can turn natural-language prompts into images, text, code, and other content, lowering the cost of producing drafts and components. Their practical impact increasingly depends on whether those pieces can become complete, dependable deliverables. This survey examines agentic artifact creation, which we define as stateful construction in which an AI system materially constructs or revises a deliverable and intermediate observations redirect later work. Functionally, the process links an operational representation of the artifact, a construction policy, and runtime verification whose feedback can redirect later actions. We reviewed 259 works available through August 20, 2026: 230 systems meeting this definition and 29 benchmarks of agentic artifact construction. We compare six artifact families, then analyze application settings and evaluation practice as separate dimensions. Across families, construction challenges reflect not only modality but also how tightly decisions are coupled and whether failures become visible while they remain repairable. Decomposition can reduce local complexity while increasing coordination and reassembly costs. Learned judges may add little independent evidence when they share the generator's preferences or blind spots. We formulate principles for keeping commitments and responsibility explicit, turning feedback into targeted repair, and revalidating affected state after change. We also identify opportunities for sustaining coherent, accountable control as artifacts, creator intent, and construction systems evolve. A curated paper list is available at https://github.com/GeminiLight/awesome-agentic-artifact-creation.

1 Introduction

The survey defines agentic artifact creation as stateful, feedback-directed construction of complete deliverables, organized through artifact state, construction policy, and runtime verification. It reviews 230 systems and 29 benchmarks to compare artifact families, applications, evaluation, and principles for inspectable control.

1.1 From Direct Generation to Agentic Creation

  • Agentic artifact creation addresses deliverables whose interdependent requirements and partially observable validity make direct generation and broad regeneration unreliable.Intermediate observations must localize failures and guide feasible repairs before delivery, as illustrated by software agents editing repository state and interpreting execution feedback [8].Source: §1.1 · pp. 1–2

1.2 Functional Overview of Agentic Creation

  • Construction couples an Operational Representation, Construction Policy, and Runtime Verification so observations redirect later decisions using persistent artifact state and feedback.Shared construction units and records support composability, traceability, and revisability by enabling targeted edits and dependency rechecks without discarding accepted work.Source: §1.2 · p. 2, Figure 1 · p. 2

1.3 Review Scope and Method

  • The review retains systems that materially construct or revise deliverables, including mixed-initiative systems, and identifies 230 systems and 29 construction benchmarks through a multi-source search and screening process.Generic problem-solving agents, observation-independent workflows, direct generators, and artifact-consuming or evaluation-only systems were excluded.Source: §1.3 · pp. 2–3

1.4 Our Contributions and Organization

  • The survey contributes a functional characterization, a coded corpus comparing six artifact families, independent analyses of applications and evaluation, and four principles for inspectable construction control.It also uses the exposed problems to structure a research agenda.Source: §1.4 · p. 3

2 Background

Agentic artifact creation combines foundation-model content operations with agentic control, artifact-production precedents, and an artifact-centered lens for comparing acceptance criteria, editable state, observations, and repair across domains.

2 Background

  • The background review progresses from generative and agentic foundations through artifact families, applications, evaluation, principles, and opportunities.Figure 3 previews this survey structure.Source: §2 · p. 4

2.1 Generative Foundation Models

  • Foundation models provide reusable operations across language, code, images, video, audio, and 3D, but do not maintain construction state, choose subsequent actions, incorporate observations, or determine completion [19].These limitations motivate adding agentic control mechanisms.Source: §2.1 · p. 4

2.2 Agentic AI Architectures

  • Agentic architectures add planning, memory, tool use, observation, and stopping, but artifact creation specifically requires preserving artifact state and carrying observations into later artifact actions or stopping [20] [21] [23] [10] [12] [13] [14].The relevant unit is the delivered artifact and its acceptance criteria, not the number of agents; human authority remains consequential.Source: §2.2 · p. 4

2.3 Artifact Production Processes

  • Earlier artifact-production research links specifications, structured action spaces, partial artifacts, feedback, and human judgment to later local edits and iterative production decisions [26].These precedents include procedural content generation, synthesis, shape grammars, and mixed-initiative co-creation, while leaving open which relations transfer across artifact families.Source: §2.3 · p. 4

2.4 Positioning Against Adjacent Surveys

  • Adjacent surveys emphasize architectures, models, modalities, creative refinement, software construction, trajectories, human control, evaluation, learned behavior, or world models, whereas this survey organizes comparison around the Delivered Artifact [19] [20] [21] [23] [32] [35].This organizing unit makes acceptance criteria, editable state, observations, and repair scope comparable across artifact families.Source: §2.4 · p. 5, Table 1 · p. 5

3 The Agentic Creation Paradigm

Agentic Artifact Creation is stateful artifact construction in which represented artifact or process state and intermediate observations guide later decisions. Its architecture links operational representation, construction policy, and runtime verification, yielding composability, traceability, and revisability when these elements align.

3 The Agentic Creation Paradigm

  • Agentic Artifact Creation treats the deliverable as inspectable, changeable state that guides later construction decisions, regardless of model family or call count.Source: §3 · p. 5

3.1 Conceptual Definition

  • An episode qualifies when the system materially constructs or revises an artifact, carries state across decisions, and uses an intermediate observation to redirect subsequent work.Redirection may change the next action, revision target, active branch, or stopping decision.Source: §3.1 · p. 5

3.2 Functional Architecture

  • The functional architecture couples state updates, policy decisions, and verification: actions update represented state, policy selects subsequent work, and feedback triggers acceptance, revision, escalation, or termination.The roles recur until acceptance or termination, without prescribing a particular implementation.Source: Figure 4 · p. 6, §3.2 · p. 6

3.2.1 Operational Representation

  • Operational Representation exposes an Intermediate Form and Edit Interface, with feasible repair depending on alignment between representational granularity, editable targets, and their dependencies.The authoritative revision substrate may be an artifact instance, structured model, or executable program, while edits may target units, relations, or the whole artifact.Source: §3.2.1 · pp. 6–7

3.2.2 Construction Policy

  • Construction Policy converts task specifications, represented state, and feedback into actions through workflow or autonomous control, while agent topology determines how responsible agents are organized.Decomposition can reduce local complexity and expose specialist interfaces, but introduces coordination costs, dependency-bearing handoffs, and questions about global commitments and surviving state.Source: §3.2.2 · p. 7

3.2.3 Runtime Verification

  • Runtime Verification derives control-relevant feedback from observations, but autonomy is limited when representation, interfaces, evidence, and policy responses are misaligned.Observation availability does not establish validity; effective feedback must support feasible responses while implicated state or decisions remain relevant.Source: §3.2.3 · p. 8

3.3 Construction Affordances

  • Aligned plans, records, interfaces, and runtime evidence provide composability, traceability, and revisability, although direct generation remains appropriate for simple artifacts whose failures are cheap to detect and regenerate.These affordances support component replacement, failure localization, targeted edits, and revalidation of affected dependencies.Source: §3.3 · p. 8

4 Landscape of Artifact Families · 4.1 Textual Artifacts

The survey organizes agentic artifact creation into six families by acceptance criteria, dependency regimes, and where consequential failures become observable. Within textual artifacts, repairability depends on representations that connect evolving commitments to targeted edits and feedback.

4 Landscape of Artifact Families

  • The six artifact families are classified by independently accepted deliverable form, with analysis centered on maintained state, available repairs, and observations that redirect construction.The families are textual, 2D visual, audio, video, spatial, and behavioral; Figure 5 contrasts static inspection, temporal playback, and runtime consequences as dominant observation points.Source: §4 · p. 8, Figure 5 · p. 9
  • Table 2 summarizes family-level dependency regimes and observation modes, while Table 3 and Figures 5–6 organize representative systems and six families into analytical profiles.These visual summaries describe tendencies and implementation categories rather than exclusive modes or unseen performance rankings.Source: §4 · p. 9, Figure 5 · p. 9, Figure 6 · p. 10

4.1 Textual Artifacts

  • Textual artifacts are judged by linguistic content, discourse organization, and communicative fit, using representations such as outlines, evidence links, memory, and revision histories.Source and schema checks provide early feedback, while reader, creator, or domain-expert responses arrive later and address different criteria.Source: §4.1 · p. 9

4.1.1 Creative Writing

  • Creative writing requires maintaining expressive commitments that extend beyond source coverage or task compliance across story worlds, voice, timing, form, and anticipated reception.Narrative systems use role separation, persistent character or social state, personas, and queryable narrative memory; performative systems retain timing, setup–punchline structure, callbacks, or performance cues, while preference feedback rarely identifies the responsible commitment.Source: §4.1.1 · p. 9

Summary

  • Repair remains bounded when feedback can be traced to the responsible event, claim, source, rule, or argument rather than relying only on fluency or preference scores.Such traceability connects observed failures to an appropriate repair unit while acknowledging that several textual qualities remain only partly observable during construction.Source: §4.1.1 · p. 9, Summary · p. 13

4.1.2 Professional Documents

  • Professional documents are accepted against source, domain, audience, or workflow obligations, with repair linked to evidence, rules, executable state, or repository structure.Informational reports emphasize coverage, traceability, and claim–source links, while functional-document systems retain interaction-derived rules, rendered HTML, code, workbook state, critiques, or paired translation state for targeted revision.Source: §4.1.2 · p. 11

4.1.3 Scholarly Manuscripts

  • Scholarly manuscripts align contribution-level claims with methods, results, and prior work, enabling revision at the claim, argument, section, or claim–evidence level.PaperDebugger, The AI Scientist, MLR-Bench, SurGE, DIAGPaper, and MARG support targeted or diagnostic revision, while research-design changes belong to the broader scientific workflow in §5.5.Source: §4.1.3 · p. 11

Summary

  • Across textual subtypes, control becomes harder as commitments extend beyond the edited passage into story worlds, external evidence, workflow obligations, or contribution-level arguments.Representations make repair more specific when they trace an inconsistency to its responsible event, relationship, source, rule, or argument, but reader and domain judgment remains necessary for coherence, novelty, persuasion, and professional fitness.Source: Summary · pp. 11, 13

4.2 2D Visual Artifacts

2D visual artifact creation uses rendered surfaces as feedback while varying the structure that authorizes those surfaces, from data transformations and visual objects to source-backed pages and deck-wide commitments. Across profiles, dependable repair requires preserving links between visible defects and the underlying data, source, semantic, and cross-page state.

4.2 2D Visual Artifacts

  • The 2D visual family spans raster layers, vectors, chart code, layout trees, and deck structures, with rendered feedback authorizing artifacts through recorded data, perceptual or semantic specifications, or organized source content.DrawAI illustrates the shared challenge of recovering editable structure from flattened images across figures, presentations, posters, and diagrams.Source: §4.2 · p. 13

4.2.1 Data Visualizations

  • Data visualizations must communicate effectively while keeping every mark faithful to source data and recorded transformations, making targeted repair depend on links among marks, encodings, transformations, and data.Executable transformation paths, synchronized chart state, and directly addressable SVG marks support revision, but code can make visual edits indirect and SVG can detach marks from their data.Source: §4.2.1 · p. 13

4.2.2 Illustrative Graphics

  • Illustrative graphics require editable structure beneath rendered images or diagrams so feedback can target perceptual regions, semantic entities, and relations rather than only the visible surface.Diagrams additionally need synchronized semantic topology, visual styling, spatial layout, and visual observation; engineering schematics impose constraints that rendering alone cannot establish.Source: §4.2.2 · pp. 13–14

4.2.3 Visual Documents

  • Visual documents must coordinate source fidelity, page hierarchy, legibility, narrative order, terminology, style, and cross-slide consistency, because local content or layout fixes can invalidate broader page or deck commitments.Poster repair links source claims, content units, page regions, and rendering checks, while presentation state carries claims, terminology, and design changes beyond the slide where feedback appears.Source: §4.2.3 · p. 14

Summary

  • Valid local renders do not prove that larger dependencies survived an edit, so bounded repair must preserve links from visible defects to data, source content, semantic objects, and cross-page commitments.Direct surface edits provide precise control but can sever underlying links, whereas structured code and object state preserve them while making some visual changes indirect.Source: Summary · p. 14

4.3 Audio Artifacts

Audio artifact construction links editable symbolic or production state to rendered sound, while different profiles validate musical structure, source–script–speaker alignment, or scene-level event alignment. Structural checks can redirect upstream edits, but perceptual qualities require listening and remain repairable only when tied to editable units.

4.3 Audio Artifacts

  • Audio artifacts are time-organized symbolic or rendered sound whose construction revises notes, phrases, synthesis, effects, mixes, or scene events against stage-specific feedback.Validation differs across profiles: musical relations for music, source–script–speaker links for spoken audio, and scene-level event alignment for soundscapes.Source: §4.3 · p. 14

4.3.1 Music

  • Music systems preserve compatibility among note-, phrase-, and lyric-level decisions, long-range harmony, form, performance intent, and production dependencies through modular agents and editable representations.Symbolic checks can redirect structural repair early, whereas timbre and expression still require listening after rendering.Source: §4.3.1 · pp. 14–15

4.3.2 Spoken Audio

  • No dedicated spoken-audio construction system appears in the mapped full-publication corpus; modular text- or video-to-audio and podcast pipelines support targeted revision only while source, script, voice, and clip links persist.Without those links, late listening feedback tends to trigger whole-segment regeneration; AudioToolAgent routes audio tools but does not maintain an editable audio artifact.Source: §4.3.2 · p. 15

4.3.3 Soundscapes

  • Soundscape systems represent acoustic events, sources, timing, edit programs, rendered clips, and scene state so omissions, misalignment, and mixing problems can become local repair units.They coordinate segmentation, retrieval, synthesis, executable plans, specialist generation, post-production, and feedback-driven refinement across complex scenes.Source: §4.3.3 · p. 15

Summary

  • Audio construction separates inspectable symbolic or production state from qualities heard after rendering, enabling structural redirection and local perceptual repair when feedback remains linked to editable units.Music links notes, phrases, harmony, and form; spoken audio links source units, scripts, voices, and segments; soundscapes link events, sources, edit programs, and rendered clips.Source: Summary · p. 15

4.4 Video Artifacts

Video artifact construction must preserve communicative obligations across shots, channels, and time by linking represented state to targeted repair. Across expository, narrative, and promotional profiles, the central bottleneck is delayed, weakly localized temporal feedback.

4.4 Video Artifacts

  • Video artifacts are ordered visual or audiovisual sequences whose acceptance criteria extend across shots and time, requiring state that supports revision before assembled consequences become visible.State may include scripts, storyboards, shot lists, timelines, event graphs, and cross-shot memory; profiles are organized by communicative purpose.Source: §4.4 · p. 15

4.4.1 Expository Videos

  • Expository videos must accurately and clearly communicate externally grounded material across assembled visuals, narration, subtitles, and timing.Systems align scripts and scenes, compile executable plans, inspect renders, and trace repair units such as claim–segment links or timeline intervals to source documents.Source: §4.4.1 · p. 15

4.4.2 Narrative Videos

  • Narrative videos must preserve story intent, identity, causal order, and cinematic decisions across shots, using hierarchical plans, externalized dependencies, and targeted revision.Event graphs and cross-shot memory support repairs to events, entities, shots, or relations, but formal checks cover only encoded constraints and prompt-level repair leaves other dependencies unenforced.Source: §4.4.2 · p. 16

4.4.3 Promotional Videos

  • Promotional videos depend on message fidelity, recognizable identity, audiovisual appeal, and audience or channel fit, so local polish can still undermine global presentation.Represented state may combine briefs, profiles, factual sources, assets, scene plans, and editable timelines; repairs can target claims, placements, shots, beats, or style decisions.Source: §4.4.3 · p. 16

Summary

  • Delayed, weakly localized temporal feedback is the shared bottleneck, because playback enables targeted repair only when failures map to responsible claims, shots, segments, transitions, or motion units.Otherwise, feedback triggers broad regeneration; represented state binds sourced claims, identities, events, and audience-facing messages to assembled channels.Source: Summary · p. 16

4.5 Spatial Artifacts

Spatial artifacts remain dependable only when geometry, relations, constraints, and execution-relevant state survive revision. Rendering, traversal, simulation, and fabrication feedback expose different failures, while local repair requires tracing observations to responsible artifact elements.

4.5 Spatial Artifacts

  • Spatial artifacts span persistent assets and scenes whose acceptance criteria strengthen from geometry and appearance to engineering, simulation, and fabrication requirements.Their state may include meshes, scene graphs, asset libraries, engine scenes, parametric histories, or CAD programs, with feedback arriving through rendering, geometry operations, traversal, simulation, or fabrication checks.Source: §4.5 · p. 16

4.5.1 3D Assets

  • 3D assets preserve semantic parts, appearance controls, edit handles, dimensions, features, constraints, and boundary geometry, but revision can break topology, design intent, solid validity, or manufacturability.Construction histories and cross-backend graphs enable targeted repair when links identify affected geometry and invalidated checks; otherwise revision may require whole-asset regeneration.Source: §4.5.1 · pp. 16–17

4.5.2 3D Scenes

  • 3D scenes must preserve multi-object relations, layout, traversal, task compatibility, component interfaces, and dimensional constraints across authoring, assembly, and use.Explicit graphs, executable code, rendered verification, solver feedback, and deterministic assembly tests expose consequences, but render quality alone does not establish collision-free, deployable behavior.Source: §4.5.2 · p. 17

Summary

  • Rendering exposes appearance, whereas traversal, simulation, and fabrication reveal failures missed by static inspection; local repair is possible only when feedback traces to a responsible part, relation, feature, constraint, or asset.The represented state expands from semantic parts and editable geometry to parameters, constraints, layouts, traversal, and component relations.Source: Summary · p. 17

4.6 Behavioral Artifacts

Behavioral artifacts are accepted through state-dependent responses, with software judged by executable or interactive behavior and simulations by fidelity to specified or external processes. Across both branches, runtime observations can redirect repair, but longer horizons broaden coverage while making failures costlier to localize.

4.6 Behavioral Artifacts

  • Behavioral artifacts are accepted through responses to inputs, actions, and environmental interaction, comprising software systems and simulation models with operational representations that guide later revisions.Family assignment follows the delivered artifact and its defining repair unit.Source: §4.6 · pp. 17–18

4.6.1 Software Systems

  • Software systems widen verification from repository execution to browser-visible action–state sequences and sampled gameplay, while longer trajectories make failures harder to localize.Repositories require consistency across implementation, dependencies, integration behavior, and executable checks; web applications synchronize source, rendered regions, and interaction; games require sampled playthroughs beyond code and assets.Source: §4.6.1 · pp. 18–19

4.6.2 Simulation Models

  • Simulation models are accepted by fidelity to a specified virtual world or an external physical process, with repair targeting state transitions, parameters, constraints, or executable model components.Virtual-world simulators require legal, reproducible transitions and consistent persistent state, whereas physical-world models must agree with observations, domain constraints, or governing laws over checked regimes.Source: §4.6.2 · p. 19

Summary

  • Across behavioral artifacts, execution reveals consequential behavior but covers only part of the relevant state space, so repair requires tracing failures to bounded editable units and revalidating affected conditions.As behavioral horizons grow, representative observations become more expensive and failed trajectories point less directly to their causes.Source: Summary · p. 19

5 Applications

Applications differ in domain objectives, artifact bundles, acceptance evidence, and human authority, while workflows may span multiple artifact families or independently accepted deliverables. Across six contexts, feedback redirects construction, but responsibility for consequential interpretation and acceptance remains with authorized humans.

5 Applications

  • Applications are organized into six contexts whose objectives, artifact bundles, tools, acceptance evidence, and human authority differ.Artifact profiles track one accepted deliverable, whereas application profiles track workflow decisions that may propagate across artifacts or stages.Source: §5 · p. 20

5.1 Creative Production

  • Creative production coordinates interdependent narrative, visual, auditory, and interactive decisions, with rendering, playback, interaction, and playability informing creator-led revision.Interactive entertainment preserves editable branching state and adds runtime validity to narrative and stylistic coherence.Source: §5.1 · p. 20

5.2 Brand Communication

  • Brand communication maintains shared identity and message commitments across campaign assets and customer interactions, using rendering checks and response signals while brand owners retain release authority.Click-through and other audience responses can redirect later work but remain noisy proxies for commercial goals.Source: §5.2 · p. 21

5.3 Educational Support

  • Educational support adapts instructional resources and learning trajectories to goals, assessments, and learner responses, while teachers retain authority over learning commitments.Maintained learner state can condition later tutoring, question generation, pacing, and practice-scenario selection without silently changing those commitments.Source: §5.3 · p. 21

5.4 Professional Work

  • Professional work transforms changing evidence and operational state into decision-relevant outputs, but qualified professionals, organizational decision-makers, or editors remain responsible for consequential interpretation and release.Provenance and passing checks can support auditability, while public information emphasizes sourcing, correction, and editorial responsibility.Source: §5.4 · p. 21

5.5 Scientific Research

  • Scientific research uses experiments, executable checks, and review either to generate new evidence or to communicate established evidence faithfully, with researchers responsible for methods, ethics, and claims.Discovery workflows revise hypotheses and analyses from observations, whereas research communication preserves source fidelity, consistency, traceability, and reviewable interpretation.Source: §5.5 · p. 22

5.6 Engineering Design

  • Engineering design validates deployable systems, fabricable targets, or simulators through execution, fabrication, simulation, or deployment, with observed consequences redirecting design or model state.Application boundaries include cross-output dependencies and constructed environments, whose downstream training roles add tasks, trajectories, rewards, datasets, and policies without replacing original acceptance criteria.Source: §5.6 · pp. 22–23

6 Evaluation and Benchmarking

Evaluation of agentic artifact creation must distinguish delivered artifacts, construction trajectories, and system properties, then align each claim with explicit signals and protocols. The review finds task-stratified evidence is essential, while trajectory, contextual-use, and beyond-capability system evaluation remain comparatively underdeveloped.

6.1 Evaluation Targets

  • Evaluation separates delivered-artifact requirements and use, construction-trajectory detection and repair, and system properties beyond task capability, each requiring distinct metrics and contrasts.Aggregate artifact or trajectory measurements support task-capability claims, whereas system-property claims require repeated runs, controlled conditions, interventions, resource tradeoffs, or version comparisons.Source: §6.1 · pp. 23–24

6.2 Evaluation Evidence

  • Evidence strength depends on pairing an observation channel with an evaluator for a stated requirement and stage, using complementary signals because multiple judges may share blind spots.Channels include reference inputs, structured state, rendered output, runtime behavior, construction history, and use outcomes; adequacy separately depends on coverage, validity, stability, and calibration.Source: §6.2 · pp. 24–25

6.3 Evaluation Protocols

  • Comparable evaluation requires fixed task sets, metrics, run configurations, and analysis plans; missing fields remain evidence gaps, and final-artifact scores cannot establish trajectory properties.Trajectory claims require records of checks, revisions, intervention, and stopping, while system claims require observations matched to the stated dimension.Source: §6.3 · p. 25

6.4 Benchmark Landscape

  • Kimi K3 leads all five Models Arena tasks, but Agents Arena performance varies from 54.3rd percentile on native Android to 87.5th on slides, showing aggregate ranks conceal task differences.Claude Fable 5 reaches the top percentile on native Android, while GPT-5.6 Sol ranges from the 87.5th percentile on slides to the 34.2nd percentile on mobile apps.Source: §6.4 · p. 26, Figure 9 · p. 26
  • Benchmarks are strongest for delivered artifacts and bounded executions; only three protocols assess system properties beyond task capability, while audio, video, and spatial benchmarks remain less mature.Reporting inconsistencies in calibration, disagreement, tools, budgets, stopping, and cost further limit comparison.Source: §6.4 · p. 26

7 Principles of Agentic Creation

The section formulates four qualitative principles for inspectable control in agentic artifact construction: externalize commitments, define responsibility, make feedback repairable, and revalidate affected state. Together, they preserve consequential commitments while allowing artifact- and application-specific processes rather than prescribing one architecture or pipeline.

7 Principles of Agentic Creation

  • Four principles connect acceptance criteria to addressable state, dependencies to responsibility, evidence to bounded repair, and change to selective revalidation.They synthesize decision interdependence, failure observability, and repairability as qualitative conditions, not independent causal factors or a universal ranking.Source: §7 · p. 27, Figure 10 · p. 28

7.1 Externalize Commitments

  • Externalize commitments by linking consequential requirements and prior decisions to addressable artifact state that later actions and checks can preserve, change, or verify.Relevant state may include acceptance criteria, constraints, semantic edit units, dependencies, provenance, or review status; representations should be sufficient rather than maximal.Source: §7.1 · p. 28

7.2 Define Control Boundaries

  • Define control boundaries where dependencies or consequences change, specifying who may act and who remains responsible for preserving commitments across handoffs.Decomposition fails when shared constraints or cross-boundary ownership are lost, and added boundaries can increase communication, rework, latency, and propagation costs.Source: §7.2 · pp. 28–29

7.3 Make Feedback Actionable

  • Make feedback actionable by connecting acceptance criteria to timely evidence, localized diagnosis, affected state, and a feasible repair or escalation.Global scores can support selection or stopping but cannot identify local repairs; repair requires an edit interface whose scope exposes the implicated failure.Source: §7.3 · p. 29

7.4 Revalidate Affected State

  • Revalidate affected state by invalidating dependent evidence and approvals when sources, requirements, tools, artifacts, or authorizations change, then selectively rechecking what remains affected.A versioned dependency and evidence graph could support invalidation, regression, rollback, and renewed review, but cross-family evidence is currently too sparse to establish one mechanism.Source: §7.4 · p. 29

8 Challenges and Opportunities

The section identifies six unresolved control problems in agentic artifact creation: coherence, targeted repair, durable evolution, evolving creator intent, delegated authority, and evaluation under multiple valid outcomes. It proposes explicit representations, targeted verification, versioned change, scoped memory, authority-aware delegation, and budget-matched evaluation as hypotheses for addressing them.

8 Challenges and Opportunities

  • Together, these challenges concern coherence, repair, system evolution, creator intent, authority, and evaluation, and proposed directions remain hypotheses requiring matched model, tool, human, and resource budgets.The principles and difficulty lens expose unresolved control problems rather than settled comparative conclusions.Source: §8 · p. 29

8.1 Global Coherence under Decision Interdependence

  • Agentic systems must preserve artifact-level commitments across decomposed stages by exposing consequential dependencies in an operational representation tied to rendered or executed behavior.Selective explicitness can materialize only high-risk or change-relevant links among requirements, artifact units, decisions, owners, and evidence.Source: §8.1 · pp. 29–30

8.2 Targeted Repair under Self-Diagnosis Limits

  • Reliable repair requires a non-correlated chain from detection through diagnosis to the smallest safe edit, with unresolved causes escalating and evaluation reporting detection, localization, repair, regression, and recovery separately.Executable checks, independent critics, expert review, and dependency-aware fault localization can diversify evidence and map failures to source state.Source: §8.2 · p. 30

8.3 System Self-Evolution beyond Task-Specific Gains

  • Durable self-evolution requires versioned patches with provenance, scope, rollback, held-out shadow evaluation, independent authorization, and attribution separating retained changes from retrieval, task order, extra inference, and evaluator bias [359] [360].Only gains that transfer, persist, avoid regressions, and are attributable to the retained patch should count as durable improvement.Source: §8.3 · p. 30

8.4 Personalization under Evolving Creator Intent

  • Personalization should use inspectable, scoped preference memory that distinguishes stable preferences, project commitments, provisional choices, and uncertain inferences while supporting withdrawal, clarification, and correction.Preferences learned in one project should not silently govern another, and new feedback should not automatically override retained intent.Source: §8.4 · pp. 30–31

8.5 Human Authority under Agentic Delegation

  • Delegation must represent authority at the decision and action level, separating evidence of correctness from permission and using ownership-based review, approval gates, reversibility, and action-level provenance.Confidence cannot substitute for authority, although controls should avoid turning routine creation into constant interruption.Source: §8.5 · p. 31

8.6 Reliable Evaluation of Open-Ended Artifacts

  • Open-ended artifacts require versioned acceptance specifications that preserve multiple valid outcomes, criterion-level results, evaluator roles, evolving preferences, and stakeholder disagreement rather than collapsing them into one score.Comparisons should match models, tools, feedback, compute, and human effort, while evaluating construction trajectories, recovery, evaluator stability, and total resource use.Source: §8.6 · p. 31

9 Conclusion

Agentic artifact creation is stateful construction in which observations of an evolving artifact redirect later work through aligned operational representation, construction policy, and runtime verification. Its value depends on whether this control enables reliable, repairable revision and state preservation at a cost justified by improved later decisions.

Definition

  • Agentic artifact creation links operational representation, construction policy, and runtime verification so evidence can localize failures and direct feasible edits.The process is defined by observations of an evolving artifact redirecting later work.Source: §9 · p. 31

Challenges and principles

  • Across artifact families, tightly coupled decisions, late or opaque failures, and coarse edits make construction challenges extend beyond modality.The four principles respond by externalizing commitments, defining control boundaries, linking feedback to feasible repairs, and revalidating affected state.Source: §9 · p. 31

When agentic construction helps

  • Direct generation is preferable when little accepted state must be preserved or agentic control costs more than regeneration.Agentic construction is worthwhile when observations improve later decisions enough to justify its control cost, with bounded revision and state preservation as key benefits.Source: §9 · p. 31

Evaluation and open challenges

  • Evaluation should match model, tool, compute, feedback, and human budgets while pairing delivered-artifact results with trajectory evidence relevant to the claim.Open challenges include preserving distant dependencies during revision, repairing undiagnosed flaws, distinguishing durable improvement from local gain, learning creator preferences without overcommitting to transient feedback, delegating with explicit authority, and judging open-ended outcomes beyond a single score.Source: §9 · p. 31

Related papers

Loading 2608.28122v1…