Agentic Systems: A Guide to Transforming Industries with Vertical AI Agents

Fouad Bousetouane

REVIEW 4 major objections 4 minor 3 cited by

Agentic Systems: A Guide to Transforming Industries with Vertical AI Agents

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A proposed 'Cognitive Skills' module aims to make vertical AI agents precise and reliable by bridging general LLM reasoning and domain-specific inference.

desk verdict A clearly written practitioner taxonomy with one genuinely new named module, but the module's core benefit is asserted, not demonstrated; useful as a design guide, thin as a research contribution. read the letter →

arxiv 2501.00881 v1 pith:N6CGDVIW submitted 2025-01-01 cs.MA

Fouad Bousetouane This is my paper · ORCID

classification cs.MA

keywords verticalAIagentsLLMcognitiveskillsagenticsystemsmulti-agenthuman-in-the-loopretrieval-augmentedgenerationdesignpatterns

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that vertical AI agents—agents built for a specific industry—should include a fourth core module, which it calls 'Cognitive Skills': a hub of purpose-built inference models tailored to domain tasks. The claim is that general-purpose or fine-tuned LLMs alone, even with tools, lack the precision for specialized tasks such as risk assessment, compliance monitoring, OCR, and guardrails. By adding this module, agents become both adaptable and precise, bridging the LLM's general reasoning with domain-specific inference. The paper also offers a taxonomy of agentic systems—task-specific, multi-agent, and human-augmented—with architectural patterns for each. A sympathetic reader would take this as a proposal for standardizing vertical AI agent design, not an experimentally proven result.

What carries the argument

The central object is the Cognitive Skills module: a collection of purpose-built inference models that sit alongside the LLM's reasoning engine and external tools. It is described as a 'bridge' between general reasoning and domain-specific inference. In the paper's architecture, the Reasoning Engine (the LLM) orchestrates inputs from Memory, Cognitive Skills, and Tools; the Cognitive Skills module supplies high-precision outputs for specialized tasks, while Tools handle knowledge retrieval and environment interaction. The paper also uses a three-category taxonomy and specific patterns (RAG Agent Router, orchestrated multi-agent, HITL) as design patterns to organize the space.

What would settle it

A concrete test would be to take a vertical domain task, say contract risk assessment or medical image screening, and compare an agent built from a fine-tuned LLM plus tools against the same agent augmented with a Cognitive Skills module containing a purpose-built model for that task, measuring task-specific accuracy on a fixed benchmark. If the fine-tuned-plus-tools agent already performs as well, the module's premise fails.

Watch

Extended reading notes

Core claim

The central discovery is a design claim: the authors propose that the core building blocks of an LLM agent should be expanded from memory, reasoning, and tools to also include a Cognitive Skills module. This module acts as a model hub containing purpose-built models specialized for tasks that general LLMs and even fine-tuned LLMs struggle with, such as credit underwriting risk assessment, vulnerability detection, compliance monitoring, OCR, image classification, speech processing, and responsible-AI guardrails. The paper states that fine-tuned models often lack the precision and specialization required for complex, domain-specific tasks, and that the Cognitive Skills module bridges the gap between the LLM's general reasoning abilities and domain-specific inference tasks. It further advances a three-category taxonomy of agentic systems (task-specific, multi-agent, human-augmented) and illustrates implementation patterns including a RAG Agent Router, a RAG Orchestrated Multi-Agent System, and a Human-in-the-Loop pattern.

Load-bearing premise

The argument rests on the claim, made without cited evidence, that fine-tuned LLMs even with tools cannot deliver the precision needed for complex domain-specific tasks; if a fine-tuned model with tool access already achieves that precision, the Cognitive Skills module is redundant.

Editorial extensions

If this is right

If the Cognitive Skills module is adopted, vertical AI agents will standardize on four core modules—Memory, Reasoning Engine, Cognitive Skills, Tools—making architectures more comparable across vendors.
Specialized inference models like risk classifiers and guardrails become first-class components rather than ad-hoc add-ons, which could improve reliability and auditability of agent outputs.
The three-category taxonomy provides a common language for describing agentic systems, which could accelerate adoption and reduce design ambiguity.
The RAG Agent Router and orchestrated multi-agent patterns offer concrete starting points for organizations building domain-specific retrieval systems.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper does not compare the Cognitive Skills module against a strong baseline; a natural extension is to test whether the module outperforms fine-tuning alone on a suite of vertical benchmarks.
The distinction between Cognitive Skills and Tools is conceptual rather than operational—both are external model calls—so a stricter boundary (e.g., inference-only vs. retrieval/action) would make the proposal easier to evaluate.
The taxonomy excludes self-improving or learning agents; a future revision might add a fourth category to account for agents that update their own skills from feedback.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

What you should know: this is a competent, readable practitioner survey whose only new idea is naming a fourth building block, the Cognitive Skills module. The rest is a sensible review of familiar patterns — ReAct, RAG, multi-agent orchestration, HITL — with concrete examples that are genuinely useful for someone building vertical agents. The prose is clean and the taxonomy (task-specific, multi-agent, human-augmented; memory, reasoning, cognitive skills, tools) is plausible as a shared vocabulary. The RAG router and orchestrated multi-agent walk-throughs are well done and would help a team starting out.

The soft spots are real but not fatal if the paper is positioned correctly. The central claim — that Cognitive Skills is a new core building block that 'fills the gap' and 'ensures' precision and reliability — is asserted, not demonstrated. Section 3.3.3 says fine-tuned models 'often lack the precision and specialization required for complex, domain-specific tasks,' but Section 2.1.1 credits vertical agents with 'exceptional precision' precisely because their reasoning engines are fine-tuned. That is a direct tension, and the paper never resolves it. There is no citation, benchmark, or ablation. The boundary with the Tools module is also fuzzy: OCR, classifiers, and guardrails are callable models, just like APIs and retrieval systems. If a fine-tuned LLM with tool access already does the job, the new module is a relabeling, not a contribution. The conclusion also overreaches with 'paradigm shift' and 'revolutionize industries.'

The stress-test note is right on all of this. I'd add that the paper's own internal logic is coherent enough — the module makes sense as a design opinion, not as an established finding. The authors should either provide a concrete example where a fine-tuned LLM plus tools fails and the skill hub succeeds, or explicitly frame the module as a recommended pattern rather than a validated requirement.

Who is this for? Practitioners and teams looking for a common vocabulary to talk about vertical agent architectures. It is not for researchers expecting empirical results. I would send it to a serious referee if the venue accepts position or survey papers, because the taxonomy is useful and the Cognitive Skills idea is worth a second opinion — but I would expect the referee to demand a major reframing and honest disclaimers. If the venue demands original results, desk reject with an invitation to resubmit as a design guide.

Referee Report

4 major / 4 minor

Summary. This position paper proposes a standardization vocabulary for Vertical AI agents by identifying four core building blocks of an LLM agent (Memory, Reasoning Engine, Cognitive Skills, Tools) and introduces the Cognitive Skills module as a new core component: a hub of purpose-built, domain-specific inference models that are claimed to bridge general LLM reasoning and specialized tasks. The paper then offers a taxonomy of agentic systems (task-specific, multi-agent, human-augmented), describes several design patterns including a RAG Agent Router and a RAG Orchestrated Multi-Agent System, and surveys industry and academic efforts. The central technical claim is that the Cognitive Skills module is necessary and novel because fine-tuned LLMs and tools alone are insufficient for precision-critical, domain-specific tasks.

Significance. If the proposed standardization and the Cognitive Skills module were rigorously supported, this could be a useful design vocabulary for practitioners building vertical agents, especially in regulated industries. The paper's strengths are its clear rhetorical organization, its concrete examples (e.g., guardrail classifiers, OCR, risk-assessment models) that illustrate what the intended module would contain, and its effort to situate the proposal within current industry frameworks (AutoGen, Semantic Kernel, Vertex AI Agent Builder). However, the manuscript contains no empirical evaluation, no formal specification, and no operationalization of the key boundary between Cognitive Skills and Tools; the significance is therefore that of an unsupported design opinion rather than a validated architecture.

major comments (4)

[§3.3.3 vs. §2.1.1] The load-bearing premise that 'Fine-tuned models often lack the precision and specialization required for complex, domain-specific tasks' is asserted without citation, benchmark, or ablation, and it directly contradicts §2.1.1, which states that vertical agents achieve 'exceptional precision' precisely by using LLMs 'fine-tuned for specialized knowledge and workflows.' If a fine-tuned LLM with tool access already delivers the required precision, the Cognitive Skills module is redundant; if it does not, the paper must provide evidence and concrete failure cases. Without such support, the central contribution is an unverified claim.
[§3.3.3 vs. §3.3.4] The manuscript never defines an operational distinction between a Cognitive Skill and a Tool. The examples of Cognitive Skills (OCR, image classifiers, guardrail classifiers, risk-assessment models) are all callable models, while the Tools module already includes knowledge retrieval, APIs, and legacy-system interfaces. The paper should specify how the Reasoning Engine invokes a Cognitive Skill differently from a Tool (e.g., different control flow, different discovery mechanism, different training/serving pipeline) and provide at least one task where a fine-tuned LLM with tool access fails but adding a purpose-built model succeeds. Without this, the claim that Cognitive Skills 'fills the gap' is not falsifiable.
[§4.3] The taxonomy of agentic systems into task-specific agents, multi-agent systems, and human-augmented agents is presented as exhaustive and primary, but no classification criterion is given and no argument is provided that these three categories cover the design space. The categories are not mutually exclusive: a task-specific agent can be a component of a multi-agent system, and human-in-the-loop can augment either a task-specific or a multi-agent system. The paper should state the intended dimension of classification (e.g., number of agents, degree of human involvement) and address overlaps, or explicitly present the taxonomy as one of several possible views rather than as an exhaustive typology.
[§3.3 and §4.2] The paper promises 'a level of standardization' and describes core modules, but it never specifies the interfaces, data flows, or control flows between Memory, Reasoning Engine, Cognitive Skills, and Tools. Consequently, the proposed Cognitive Skills module cannot be implemented, tested, or compared against alternative designs from the manuscript alone. At minimum, the paper should provide a formal-ish definition of the module boundary, the inputs/outputs of a Cognitive Skill, and a protocol for the Reasoning Engine to select among skills; without this, the standardization claim is not actionable.

minor comments (4)

[§2.1.3] The text contains typographical inconsistencies: 'A WS' should be 'AWS' (also in §4.2.1), and 'Vertical Ai agent solutions' should be 'Vertical AI agent solutions.'
[§3.3.3] The phrase 'as aversatile, purpose-driven system' is a typo; it should read 'as a versatile, purpose-driven system.'
[References] Reference [9] (OpenAI DevDay) and [10] (AWS SageMaker) lack URLs and access dates, and reference [12] (MedAide) lacks a venue or arXiv identifier; please complete the bibliographic entries.
[Figures] Figures 1, 2, and 4 are referenced in the text but not displayed in the manuscript; please include the figures or clearly indicate that they are omitted.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is a conceptual design guide with no derivations, fitted parameters, or self-citation chain that would make a result equivalent to its inputs by construction.

full rationale

This manuscript makes no mathematical derivation and offers no empirical prediction. Its central proposal, the Cognitive Skills module, is presented as a design recommendation: the paper asserts that purpose-built inference models can improve domain-specific agent performance, but it does not fit any parameter to data and then rename that fit as a prediction. The load-bearing premise that fine-tuned LLMs often lack precision for complex domain tasks is an unsupported claim, which is a correctness or evidence concern, not circularity. No equation or construction in the paper makes one quantity equal another by definition, and the paper does not rely on its own prior results; the cited works are external frameworks and papers. The overlap between Cognitive Skills examples and the Tools module is a boundary-definition issue, not a circular reduction, since the paper never claims to derive the module's effectiveness from the Tools module or from any prior theorem. Because there is no derivation chain to walk, the circularity burden is minimal and the honest finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 1 invented entities

The paper's central proposal is a conceptual architectural module. It introduces no fitted parameters, but it relies on an unverified assumption that fine-tuned LLMs are insufficient without specialized model hubs, and it posits one new conceptual entity without independent evidence.

assumptions (3)

domain assumption Fine-tuned LLMs lack precision and specialization for complex domain tasks.
Stated in Section 3.3.3 without evidence; it is load-bearing because it motivates the need for a separate Cognitive Skills module.
domain assumption The modular decomposition into memory, reasoning, cognitive skills, and tools is both beneficial and sufficient for vertical agents.
Presented as the framework in Section 3.3 with no comparison against alternative decompositions or architectures.
domain assumption The three-category taxonomy of agentic systems (task-specific, multi-agent, human-augmented) is exhaustive and useful.
Section 4.3 introduces the categories without justification or evidence that they cover the meaningful design space.

invented entities (1)

Cognitive Skills Module
purpose: A model hub of purpose-built models (risk assessment, OCR, image classification, guardrails, etc.) that bridges LLM reasoning and external tools.
No falsifiable predictions or evaluation are provided; it is a proposed taxonomy whose benefits are asserted, not demonstrated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Agentic Systems: A Guide to Transforming Industries with Vertical AI Agents." pith.science (2026). https://pith.science/paper/N6CGDVIW

@misc{pith2026250100881,
  author       = {Pith},
  title        = {Pith review of: Agentic Systems: A Guide to Transforming Industries with Vertical AI Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N6CGDVIW}},
  note         = {Machine review of arXiv:2501.00881}
}

read the original abstract

The evolution of agentic systems represents a significant milestone in artificial intelligence and modern software systems, driven by the demand for vertical intelligence tailored to diverse industries. These systems enhance business outcomes through adaptability, learning, and interaction with dynamic environments. At the forefront of this revolution are Large Language Model (LLM) agents, which serve as the cognitive backbone of these intelligent systems. In response to the need for consistency and scalability, this work attempts to define a level of standardization for Vertical AI agent design patterns by identifying core building blocks and proposing a \textbf{Cognitive Skills } Module, which incorporates domain-specific, purpose-built inference capabilities. Building on these foundational concepts, this paper offers a comprehensive introduction to agentic systems, detailing their core components, operational patterns, and implementation strategies. It further explores practical use cases and examples across various industries, highlighting the transformative potential of LLM agents in driving industry-specific applications.

Figures

Figures reproduced from arXiv: 2501.00881 by the authors.

**Figure 2.** Example of LLM Workflow: Chain Prompting with RAG for [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Architecture of the RAG Agent Router with Domain-Specific Vec [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Architecture of the RAG Orchestrated Multi-Agent System for [PITH_FULL_IMAGE:figures/full_fig_p022_4.png]

**Figure 5.** Figure 5: Human-in-the-Loop (HITL) Agent Pattern for Collaborative [PITH_FULL_IMAGE:figures/full_fig_p026_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Generative AI for Industrial Contour Detection: A Language-Guided Vision System
cs.CV 2025-08 reject novelty 4.0 of 10

A GAN-plus-VLM pipeline improves industrial remnant contour extraction, with GPT-image-1 outperforming Gemini 2.0 Flash on SSIM, LPIPS, and Hausdorff distance.
An Agentic AI for a New Paradigm in Business Process Development
cs.AI 2025-07 reject novelty 4.0 of 10

Business processes can be modeled as goal-driven agent teams where goals, objects, and agents replace fixed task sequences, and workflows emerge from trigger objects.
Vibe Coding vs. Agentic Coding: Fundamentals and Practical Implications of Agentic AI
cs.SE 2025-05 conditional novelty 3.0 of 10

A qualitative taxonomy positions vibe coding and agentic coding as complementary paradigms rather than rivals in AI-assisted software development.

Reference graph

Works this paper leans on

15 extracted references · 10 canonical work pages · cited by 3 Pith papers

[1]

Building effective agents

Anthropic. Building effective agents. https://www.anthropic.com/ research/building-effective-agents . november , 2024

work page 2024
[2]

Ai for it operations (aiops) on cloud platforms: Reviews, opportunities and challenges

Qian Cheng, Doyen Sahoo, Amrita Saha, Wenzhuo Yang, Chenghao Liu, Gerald Woo, Manpreet Singh, Silvio Saverese, and Steven CH Hoi. Ai for it operations (aiops) on cloud platforms: Reviews, opportunities and challenges. arXiv preprint arXiv:2304.04661, 2023

arXiv 2023
[3]

Vertex ai agent builder

Google Cloud. Vertex ai agent builder. https://cloud.google.com/ products/agent-builder, 2024. Accessed: December 26, 2024

work page 2024
[4]

Agents - langchain v0.1 documentation

LangChain Developers. Agents - langchain v0.1 documentation. https: //python.langchain.com/v0.1/docs/modules/agents/, 2024. Ac- cessed: December 26, 2024

work page 2024
[5]

Magentic-one: A generalist multi-agent sys- tem for solving complex tasks

Adam Fourney, Gagan Bansal, Hussein Mozannar, Cheng Tan, Eduardo Salinas, Friederike Niedtner, Grace Proebsting, Griffin Bassman, Jack Gerrits, Jacob Alber, et al. Magentic-one: A generalist multi-agent sys- tem for solving complex tasks. arXiv preprint arXiv:2411.04468, 2024. 30

arXiv 2024
[6]

Retrieval-augmented generation for large language models: A survey

Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023

arXiv 2023
[7]

Prompt engineering with chatgpt: A guide for academic writers

Louie Giray. Prompt engineering with chatgpt: A guide for academic writers. Annals of Biomedical Engineering, 51(12):2629–2633, 2023

work page 2023
[8]

Microsoft’s agentic frameworks: Autogen and semantic ker- nel

Microsoft. Microsoft’s agentic frameworks: Autogen and semantic ker- nel. Microsoft Developer Blogs, 2023. Accessed: December 26, 2024

work page 2023

Show all 15 references

[9]

New models and developer products announced at devday,

OpenAI. New models and developer products announced at devday,
[10]

Learn how to build and deploy tool-using llm agents using aws sagemaker jumpstart foundation models, 2023

Amazon Web Services. Learn how to build and deploy tool-using llm agents using aws sagemaker jumpstart foundation models, 2023. Ac- cessed: December 26, 2024

2023
[11]

Srsa: A cost-efficient strategy-router search agent for real-world human-machine interactions

Yaqi Wang and Haipei Xu. Srsa: A cost-efficient strategy-router search agent for real-world human-machine interactions. arXiv preprint arXiv:2411.14574, 2024

2024 arXiv
[12]

Medaide: Towards an omni medical aide via specialized llm-based multi-agent col- laboration, 2024

Jinjie Wei, Dingkang Yang, Yanshu Li, Qingyao Xu, Zhaoyu Chen, Mingcheng Li, Yue Jiang, Xiaolu Hou, and Lihua Zhang. Medaide: Towards an omni medical aide via specialized llm-based multi-agent col- laboration, 2024

2024
[13]

React: Synergizing reasoning and acting in language models

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022

2022 arXiv
[14]

Kg4diagnosis: A hier- archical multi-agent llm framework with knowledge graph enhancement for medical diagnosis, 2024

Kaiwen Zuo, Yirui Jiang, Fan Mo, and Pietro Lio. Kg4diagnosis: A hier- archical multi-agent llm framework with knowledge graph enhancement for medical diagnosis, 2024. 31

2024
[2023]

Accessed: December 26, 2024

2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Building effective agents

Anthropic. Building effective agents. https://www.anthropic.com/ research/building-effective-agents . november , 2024

work page 2024

[2] [2]

Ai for it operations (aiops) on cloud platforms: Reviews, opportunities and challenges

Qian Cheng, Doyen Sahoo, Amrita Saha, Wenzhuo Yang, Chenghao Liu, Gerald Woo, Manpreet Singh, Silvio Saverese, and Steven CH Hoi. Ai for it operations (aiops) on cloud platforms: Reviews, opportunities and challenges. arXiv preprint arXiv:2304.04661, 2023

arXiv 2023

[3] [3]

Vertex ai agent builder

Google Cloud. Vertex ai agent builder. https://cloud.google.com/ products/agent-builder, 2024. Accessed: December 26, 2024

work page 2024

[4] [4]

Agents - langchain v0.1 documentation

LangChain Developers. Agents - langchain v0.1 documentation. https: //python.langchain.com/v0.1/docs/modules/agents/, 2024. Ac- cessed: December 26, 2024

work page 2024

[5] [5]

Magentic-one: A generalist multi-agent sys- tem for solving complex tasks

Adam Fourney, Gagan Bansal, Hussein Mozannar, Cheng Tan, Eduardo Salinas, Friederike Niedtner, Grace Proebsting, Griffin Bassman, Jack Gerrits, Jacob Alber, et al. Magentic-one: A generalist multi-agent sys- tem for solving complex tasks. arXiv preprint arXiv:2411.04468, 2024. 30

arXiv 2024

[6] [6]

Retrieval-augmented generation for large language models: A survey

Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023

arXiv 2023

[7] [7]

Prompt engineering with chatgpt: A guide for academic writers

Louie Giray. Prompt engineering with chatgpt: A guide for academic writers. Annals of Biomedical Engineering, 51(12):2629–2633, 2023

work page 2023

[8] [8]

Microsoft’s agentic frameworks: Autogen and semantic ker- nel

Microsoft. Microsoft’s agentic frameworks: Autogen and semantic ker- nel. Microsoft Developer Blogs, 2023. Accessed: December 26, 2024

work page 2023