{"id":"069d30ab-b745-4775-a6a2-3a0ed59cd511","arxiv_id":"2507.01376","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey that clarifies the overlapping concepts of AI agents, LLM agents, and Agentic AI, and maps their potential roles in smart manufacturing.","lead":"This paper reviews how AI assistants and autonomous agents could be used in factories, and explains the newer idea of 'Agentic AI'. It sorts out the vocabulary, maps capability levels, and lists technical and organizational challenges for manufacturers.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central vision rests on an unstated continuity assumption: today's retrieval/diagnostic assistants scale to oversight-free Agentic AI, but no cited system demonstrates the required independent execution.","rationale":"The reader's weakest-assumption analysis identifies the same core issue: Section 5 extrapolates from retrieval-augmented assistants to fully autonomous Agentic AI without evidence that the technology challenges in Section 6 will be resolved. My independent reading confirms this is the single most load-bearing concern. The paper is otherwise a competent background review: its historical framing in Sections 2 and 3 is standard, the cited manufacturing examples are real, and the Section 6 challenges are acknowledged in good faith. The problem is not the review's factual content but the modal strength of the forward-looking claims. Specifically, Section 5.2's 'without human oversight' is a categorical assertion that no cited system supports; Section 5.4's 'continuously improving models without human intervention' is likewise unsupported by any deployment described in the manuscript. The 'systematic review' label also lacks a documented search protocol, which compounds the issue, but the more load-bearing concern is the evidence gap for the central autonomy claim. Because the reader already recommended CONDITIONAL for precisely these reasons, my stress-test does not change the verdict; it reinforces it. The concrete test I propose would turn the concern into a checkable fact: if the cited systems all remain at the low-agenticness end, the paper should explicitly mark Section 5 as a research agenda rather than an established trajectory.","tokens_in":11196,"tokens_out":1777,"duration_ms":24921,"concrete_test":"Build a capability inventory from the three cited manufacturing implementations (Lin et al. IMVA, Jeon et al. ChatCNC, and Heredia Alvaro et al. ceramic-tile RAG) and score each against the four agenticness dimensions in Section 3.3: goal complexity, environmental complexity, adaptability, and independent execution. For each system, determine specifically whether any described action—workflow modification, schedule change, or logistics reconfiguration—occurs without human approval. If all three systems fall into 'retrieval/recommendation with human-in-the-loop' on the independent-execution dimension, then the Section 5.2 'without human oversight' claim has zero cited support, and the paper's language should be revised to conditional or aspirational phrasing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that Agentic AI will move manufacturing from reactive task optimization to proactive, system-level, self-optimizing intelligence. For that claim to hold, the current GenAI-enabled agents described in Section 4 must be points on a trajectory that reaches high agenticness. This continuity assumption is load-bearing but unsupported. Every manufacturing implementation cited in Section 4 is an assistant or diagnostic system: IMVA performs natural-language retrieval and report generation; ChatCNC answers queries about CNC data and IIoT readings; the ceramic-tile RAG system diagnoses defects and proposes solutions. None of these is shown to autonomously change production workflows, reconfigure supply logistics, or redefine production objectives. Yet Section 5.2 states that an Agentic AI system 'can autonomously modify production workflows, identify alternative materials, and reconfigure supply logistics without human oversight,' and Section 5.4 claims systems will 'continuously improve their models ... without human intervention.' The paper itself concedes in Section 4.3 that 'practical implementations remain in the early stages.' Section 3.3 defines agenticness as a spectrum but provides no threshold or evidence that current LLM/MLLM agents are near, or even moving toward, the high-agenticness end. Thus the central claim conflates potential value with demonstrated feasibility. This is not a criticism of the review's factual background; it is a specific gap between the evidence presented and the strength of the forward-looking assertions. The review would be equally useful, and more defensible, if the Section 5 statements were explicitly framed as research hypotheses or aspirational targets rather than capabilities of existing systems.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a survey-style review of AI agents and Agentic AI in the context of future manufacturing. It traces the evolution of AI from symbolic and connectionist paradigms to LLMs and MLLMs, then presents a staged account of GenAI-enabled AI agents (LLM-Agents, MLLM-Agents) and the emerging concept of Agentic AI. The review describes applications of such agents in manufacturing, including semantic retrieval, multimodal perception, adaptive learning, and autonomous decision-making, and it identifies technical, workforce, and accountability challenges. The central claim is that Agentic AI will shift manufacturing from reactive, rule-based task optimization to proactive, system-level, self-optimizing intelligence, with capabilities such as autonomous goal formulation, adaptive planning, system-wide orchestration, and continuous learning without human intervention.","tokens_in":11549,"tokens_out":2806,"duration_ms":33112,"significance":"The paper provides a useful, clearly structured taxonomy of concepts—LLM-Agents, MLLM-Agents, and Agentic AI—grounded in traceable external definitions (notably OpenAI's agenticness framework [41] and Gartner's trend report [40]) and in recent manufacturing applications such as IMVA, ChatCNC, and the ceramic-tile RAG system. The historical background and the cited examples are largely accurate and well-documented. The review could serve as a helpful entry point for researchers entering this area, especially because it consolidates scattered terminology and connects it to manufacturing. However, the paper's forward-looking claims in Section 5 are repeatedly phrased as established capabilities rather than as a research vision or hypothesis. That phrasing, if left unchanged, overstates the evidence base and weakens the paper's scientific credibility. The survey value is real; the speculative parts need to be explicitly framed as potential, not demonstrated.","major_comments":[{"comment":"The paper states in Section 5.2 that an Agentic AI system 'can autonomously modify production workflows, identify alternative materials, and reconfigure supply logistics without human oversight' and in Section 5.4 that it can 'continuously improve its models ... without human intervention.' These are presented as current or near-term capabilities. Yet Section 4.3 itself concedes that 'practical implementations remain in the early stages,' and the manufacturing examples cited in Sections 4.1–4.2 (IMVA, ChatCNC, ceramic-tile RAG) are retrieval, diagnostic, or report-generation assistants—none is shown to autonomously change production workflows or reconfigure supply logistics. The central claim of the paper therefore conflates potential value with demonstrated feasibility. This is load-bearing. I ask the authors to either provide concrete evidence or a carefully caveated roadmap for how today's assistants scale to the described autonomy, or to rewrite Sections 5.2 and 5.4 so that these are explicitly labeled as open research opportunities and speculative boundary scenarios, not established facts.","section":"§5.2 and §5.4"},{"comment":"The definition of agenticness as a spectrum is useful, but the paper adds that 'as these capabilities reach a sufficiently high threshold, AI agents naturally transition into Agentic AI systems' without specifying any threshold, operational metric, or evidence that current LLM/MLLM-based agents are moving toward that threshold. This makes the transition claim unfalsifiable in its current form. The paper should either propose concrete measurable dimensions (e.g., task completion rate, degree of human intervention, adaptability under distribution shift) or treat the trajectory to high agenticness as an explicitly open empirical question, linked to the challenges catalogued in Section 6. Without this, the continuity assumption from retrieval assistants to autonomous agents is unsupported.","section":"§3.3"},{"comment":"The claim of 'system-level orchestration' across production, logistics, and enterprise management is stated without a single cited implementation or prototype that demonstrates cross-subsystem autonomous coordination in manufacturing. Given the paper's own admission in Section 4.3 that implementations are early-stage, Section 5.3's assertions about autonomous synchronization of scheduling, inventory, and transportation should be reframed as a design goal or research direction, or be supported by relevant literature on multi-agent orchestration in other domains (e.g., traffic or finance) that the paper could discuss as existence proofs.","section":"§5.3"}],"minor_comments":[{"comment":"The model name 'LLaV A' appears with a stray space; it should be 'LLaVA'.","section":"§3.2"},{"comment":"The citation 'Heredia ´Alvaro et al.' has a formatting issue with the surname; it should be written consistently as 'Heredia Álvaro et al.' or 'J. A. Heredia Álvaro'.","section":"§4.2"},{"comment":"The phrase 'modern manufacturing' in the abstract and introduction is used broadly; the paper would benefit from a brief scoping statement about which manufacturing sectors and process types are included (e.g., discrete vs. process manufacturing, high-mix low-volume vs. mass production).","section":"Abstract and Introduction"},{"comment":"Figure 3 presents the evolutionary path from rule-based expert systems to Agentic AI, but the figure's annotation of the 'agentic threshold' is not explained in the text; please either remove or define it in Section 3.3.","section":"Figure 3"},{"comment":"The section on cross-format document parsing is terse; adding one or two concrete examples of failure modes (e.g., scanned engineering drawings vs. text formulas) would make the challenge more tangible for a manufacturing audience.","section":"§6.1.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable survey with a solid background, but the overstatement in Section 5 is likely to be the main source of reviewer criticism. The authors should be asked to distinguish clearly between what has been demonstrated in manufacturing and what is a speculative vision. The self-citations ([2], [27]) are minor and appropriate as background. The journal may decide whether the survey scope is sufficiently novel relative to existing LLM-agent surveys; currently, the manufacturing-specific application framing is the main contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a well-meaning survey that gets the taxonomy right and the evidence blurry. The new contribution is a manufacturing-oriented mapping of LLM-Agents, MLLM-Agents, and Agentic AI, with a staged evolution story in Figures 2 and 3. Nothing here is formally new; the definitions come from OpenAI, Gartner, and existing surveys. But the synthesis is clean, and the four-component LLM-Agent architecture and the agenticness dimensions are explained accurately.\n\nWhat the paper does well: the historical narrative from expert systems to agentic AI is readable, the manufacturing examples (IMVA, ChatCNC, ceramic-tile RAG) are real and correctly described, and Section 6 lists genuine barriers like document parsing, multimodal alignment, and interpretability. The reference list is broad and mostly appropriate.\n\nWhere it wobbles: Section 5 crosses from 'could' to 'can.' Section 5.2 says an Agentic AI system 'can autonomously modify production workflows... without human oversight,' but every cited implementation in Section 4 is an assistant or diagnostic tool. Section 4.3 itself concedes 'practical implementations remain in the early stages.' Section 5.4 says systems will 'continuously improve their models... without human intervention.' That is a research hypothesis, not an established capability. The paper also calls itself a systematic review but gives no search protocol; it is a narrative review, which is fine if labeled honestly. The stress-test concern about the continuity assumption is correct: there is no basis to assume today's RAG agents lie on a trajectory to high agenticness just because agenticness is defined as a spectrum.\n\nFor a survey-oriented venue, I'd send this to review with a request for major revision: reframe Section 5 as future research directions, soften the categorical language, and either add a methodology paragraph or replace 'systematic review' with 'narrative review.' The core taxonomy is useful for practitioners and newcomers to manufacturing AI; it doesn't claim to be more than a map. I'd not cite it for any technical result, but I might cite it as a concise entry point to the manufacturing-agent vocabulary.\n\nVerdict: deserves peer review with revisions, not rejection. Clear thinking, honest limitations in Section 6, but load-bearing overstatement in Section 5.","headline":"A clear manufacturing-centric taxonomy of AI agents, undermined by Section 5 claims that treat speculative capabilities as established facts.","tokens_in":12040,"tokens_out":2604,"would_cite":false,"duration_ms":28102,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This review paper argues that Agentic AI moves manufacturing from reactive task optimization to proactive, system-level intelligence, with the potential to make manufacturing ecosystems self-optimizing and continuously evolving.","keywords":["AI Agents","Agentic AI","Generative AI","Large Language Models","Multimodal LLMs","Smart Manufacturing","Autonomous Decision-Making","Retrieval-Augmented Generation"],"falsifier":"A controlled field trial in which an agent is given authority to re-plan production, substitute materials, and reconfigure logistics through a simulated supply-chain disruption, with every human takeover logged, would settle the central transition claim, because the Agentic AI stage is defined by independent execution and the claim fails if the agent cannot close the loop without supervision in most trials.","tokens_in":1772,"feed_emoji":"🏭","tokens_out":4495,"duration_ms":110387,"temperature":0.7,"pith_summary":"This review paper organizes the overlapping concepts of AI agents, LLM-Agents, MLLM-Agents, and Agentic AI into one evolutionary narrative, and claims the narrative has a direction: from task-optimizing tools to systems that set their own manufacturing goals. The paper argues that Agentic AI signals a move from reactive task optimization to proactive system-level intelligence, emphasizing autonomy, adaptability, system-wide coordination, and continuous learning. If true, future manufacturing would shift from fixed automation to autonomous, goal-driven decision-making across production, logistics, and enterprise management, with systems that redefine objectives as conditions change and learn continuously without periodic retraining. The paper also states that such systems remain in early stages, while cataloguing open challenges in document parsing, multimodal alignment, interpretability, workforce adoption, accountability, and return on investment.","feed_headline":"Agentic AI moves manufacturing from fixed tasks to self-set goals","feed_subtitle":"Systems that set their own goals, adapt to supply shocks, and improve continuously could replace fixed production rules.","key_machinery":"The load-bearing conceptual machinery is the staged evolutionary path of agent technology, from rule-based expert systems to AI-powered decision-making to LLM-Agents to MLLM-Agents to Agentic AI, together with two definitions that give the stages content. LLM-Agents are characterized by four modules, profiling, memory, planning, and action, which the paper takes from a survey of LLM-based autonomous agents. Agenticness is defined through four dimensions, goal complexity, environmental complexity, adaptability, and independent execution, taken from an industry governance source, and the paper treats it as a gradual spectrum rather than a binary class. The bridge from today's systems to the future is retrieval-augmented generation (RAG) and knowledge graphs, which ground the paper's cited manufacturing examples, such as a semiconductor virtual assistant, a conversational CNC monitoring system, and an MLLM-based ceramic tile quality-control system, in real plant data.","core_discovery":"The central claim is that GenAI-enabled AI agents fall into LLM-Agents and MLLM-Agents, both of which remain task-oriented optimizers, while Agentic AI represents a qualitatively new stage defined by a spectrum of agenticness with four dimensions: goal complexity, environmental complexity, adaptability, and independent execution. Once these capabilities cross a threshold, the paper argues, agents transition from optimizing fixed schedules to autonomously defining, refining, and executing manufacturing goals in dynamic environments with minimal human intervention. In manufacturing this shift appears as four transitions: from task execution to goal-driven optimization, from rule-based control to adaptive planning, from localized optimization to system-level orchestration, and from static execution to continuous learning and evolution, giving Agentic AI the potential to transform manufacturing ecosystems into self-optimizing, continuously evolving systems.","pith_inferences":["The paper's evidence supports the assistant stage but not the autonomy stage: its cited systems are retrieval-augmented assistants, so the scaling from these to fully autonomous Agentic AI is an assumption rather than a demonstrated result.","The agenticness-as-spectrum definition invites a practical maturity rubric that ranks any manufacturing AI deployment along the four dimensions, telling a plant whether it is buying an assistant or an agent.","If system-level orchestration is the payoff, the natural benchmark is cross-functional: test an agent on a combined scheduling, logistics, and inventory disruption rather than on isolated tasks, because the paper's value claim is that siloed optimization underperforms orchestration.","The accountability and ROI discussion implies an economic gate left open, so a cost-benefit framework for decisions taken without human sign-off would be a direct next step."],"forward_implications":["Manufacturers would move from fixed schedules to goal-driven optimization, with an agentic system dynamically resetting production objectives for throughput, energy, and resource allocation as market and supply-chain conditions shift.","On a supply-chain disruption, the system would autonomously modify production workflows, identify alternative materials, and reconfigure logistics without human oversight.","Optimization becomes system-wide, coordinating production, logistics, and enterprise management together, which the paper argues is crucial for high-mix, low-volume manufacturing.","AI systems would improve continuously through self-supervised and reinforcement learning instead of periodic retraining, reducing waste and energy use over extended operational cycles.","Deployment is gated by open problems the paper itself lists: cross-format document parsing, multimodal alignment, black-box interpretability, workforce resistance, accountability, and unclear return on investment."],"supporting_citations":[{"why":"Supplies the survey of LLM-based autonomous agents whose four-module architecture defines the LLM-Agent stage.","marker":"[39]"},{"why":"Provides the industry governance definition of agenticness as a spectrum across goal complexity, environmental complexity, adaptability, and independent execution.","marker":"[41]"},{"why":"Gives the survey definition of Agentic AI as autonomous systems pursuing complex goals with minimal human oversight.","marker":"[15]"},{"why":"Positions Agentic AI as a top strategic technology trend, used to argue the paradigm's significance.","marker":"[40]"},{"why":"Supplies the hybrid retrieval-augmented generation mechanism for smart manufacturing that grounds the knowledge-enhanced semantic retrieval capability.","marker":"[42]"},{"why":"Presents the semiconductor intelligent manufacturing virtual assistant, the paper's working example of an LLM-Agent integrating plant systems through natural language.","marker":"[44]"},{"why":"Presents ChatCNC, a conversational machine monitoring system with real-time RAG, used as the example of real-time data retrieval for decision support.","marker":"[45]"},{"why":"Presents the RAG system for ceramic tile quality control using MLLMs, used as the example of multimodal diagnostics with retrievable domain knowledge.","marker":"[46]"}],"fun_headline_variants":["Agentic AI flips manufacturing from task-doers to goal-setters","Agentic AI: the shift from fixed schedules to self-set production goals","Four shifts: how Agentic AI turns factories into self-evolving systems","Agentic AI crosses a threshold: from task optimization to goal autonomy"],"cache_read_input_tokens":14080,"weakest_assumption_plain":"The central claim rests on assuming that today's retrieval-augmented and multimodal assistants will scale into fully autonomous, goal-formulating systems despite the unresolved technology challenges the paper lists, namely document parsing, multimodal alignment, and interpretability, because the paper's own examples only demonstrate the assistant stage.","fun_headline_variants_meta":{"raw":{"variants":["Agentic AI flips manufacturing from task-doers to goal-setters","Agentic AI: the shift from fixed schedules to self-set production goals","Four shifts: how Agentic AI turns factories into self-evolving systems","Agentic AI crosses a threshold: from task optimization to goal autonomy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000474,"raw_usage":{"total_tokens":2335,"prompt_tokens":909,"completion_tokens":1426,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":1349}},"tokens_in":525,"tokens_out":1426,"duration_ms":10393,"temperature":1.0,"reasoning_tokens":1349,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:51:47.579049+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled field trial in which an agent is given authority to re-plan production, substitute materials, and reconfigure logistics through a simulated supply-chain disruption, with every human takeover logged, would settle the central transition claim, because the Agentic AI stage is defined by independent execution and the claim fails if the agent cannot close the loop without supervision in most trials.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the industry governance definition of agenticness as a spectrum across goal complexity, environmental complexity, adaptability, and independent execution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Positions Agentic AI as a top strategic technology trend, used to argue the paradigm's significance."},{"cited_title":"Wan, Empowering llms by hybrid retrieval-augmented generation for domain-centric q&a in smart manufacturing, Advanced Engineering Informatics (2025)","cited_arxiv_id":null,"evidence_quote":"Supplies the hybrid retrieval-augmented generation mechanism for smart manufacturing that grounds the knowledge-enhanced semantic retrieval capability."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Presents ChatCNC, a conversational machine monitoring system with real-time RAG, used as the example of real-time data retrieval for decision support."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Presents the RAG system for ceramic tile quality control using MLLMs, used as the example of multimodal diagnostics with retrievable domain knowledge."}],"review_version":1}