Agentic AI Frameworks: Architectures, Protocols, and Design Challenges

Haithem Mazeni; Hana Derouiche; Zaki Brahmi

REVIEW 3 major objections 3 minor 5 cited by

Agentic AI Frameworks: Architectures, Protocols, and Design Challenges

T0 review · 3 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This review paper argues that Agentic AI frameworks can be classified by architecture, communication, memory, safety, and service-oriented alignment, and that analyzing four communication protocols reveals open challenges.

desk verdict A timely, useful survey of agent frameworks whose 'foundational taxonomy' claim hangs on sample selection that the abstract does not justify. read the letter →

arxiv 2508.10146 v1 pith:6C7WHUCQ submitted 2025-08-13 cs.AI

Hana Derouiche , Zaki Brahmi , Haithem Mazeni This is my paper

classification cs.AI

keywords AgenticAImulti-agentsystemsLLMagentsframeworkcomparisontaxonomyagentcommunicationprotocolsservice-orientedcomputingopenchallenges

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

The paper is a systematic review of seven Agentic AI frameworks and four agent-communication protocols. It tries to establish that the field can be organized into a foundational taxonomy using five axes: architecture, communication, memory, safety, and alignment with service-oriented computing. If the taxonomy holds, researchers and practitioners gain a common vocabulary for comparing agent systems and a clearer map of where scalability, robustness, and interoperability remain unsolved. The paper also positions protocol analysis as central to solving agent-to-agent communication.

What carries the argument

The carrying mechanism is the comparative evaluation design: five dimensions (architecture, communication mechanisms, memory management, safety guardrails, service-oriented alignment) applied uniformly to the seven frameworks, plus a protocol-level analysis of CNP, A2A, ANP, and Agora. These axes do the work of turning heterogeneous agent systems into commensurate categories, and the protocol analysis supplies the communication layer of the taxonomy.

What would settle it

Take any widely used agent framework not among the seven and attempt to place it in the taxonomy; if it cannot be classified without stretching a category, the taxonomy's foundational status fails. Likewise, if two frameworks that behave differently in practice land in the same cell on all five axes, the dimensions are too coarse.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes a comparative basis for Agentic AI by evaluating CrewAI, LangGraph, AutoGen, Semantic Kernel, Agno, Google ADK, and MetaGPT across five design dimensions, and by examining the Contract Net Protocol, Agent-to-Agent, Agent Network Protocol, and Agora for communication. From this comparison it derives a foundational taxonomy and identifies key limitations, emerging trends, and open challenges. Its proposed future directions target scalability, robustness, and interoperability, with service-oriented alignment as a distinguishing lens.

Load-bearing premise

The taxonomy is only as good as the sample: the paper assumes the seven frameworks and four protocols represent the Agentic AI field, with no described method for choosing them, and assumes the five evaluation axes are the right lens.

Editorial extensions

If this is right

A shared taxonomy lets practitioners compare frameworks on the same axes instead of relying on vendor descriptions.
The protocol analysis makes agent communication a first-class design dimension, pointing toward interoperability standards.
Identified limitations become a shortlist of open problems: scalability, robustness, and service-oriented interoperability.
The review can serve as a structured entry point for researchers new to Agentic AI.

Reading between the lines

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

We infer the five-axis taxonomy is measurable enough to seed a benchmark rubric: new agent frameworks could be scored per axis and assigned a category, a step the review does not itself take.
The paper leaves open whether the taxonomy covers agent systems outside the seven chosen frameworks; applying it to a broader sample would test its claim to be foundational.
The emphasis on service-oriented alignment suggests a convergence between agent frameworks and traditional service composition; a concrete next study could map each framework onto a standard service choreography.

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

Short version: this is a survey/taxonomy paper covering seven LLM agent frameworks (CrewAI, LangGraph, AutoGen, Semantic Kernel, Agno, Google ADK, MetaGPT) and four protocols (CNP, A2A, ANP, Agora). If the full text delivers what the abstract promises, it will be a genuinely useful consolidation for practitioners trying to choose between tools. It is not a new scientific result, and it does not claim to be. The most important thing to check is the sample construction: why these seven frameworks and four protocols, and what criteria were used for inclusion or exclusion.

The paper does several things well. The scope is timely and the chosen comparison axes—architecture, communication, memory, safety, and service-oriented alignment—are sensible and cover the dimensions that actually matter. The protocol coverage is a nice addition; most surveys stop at the frameworks. The abstract also explicitly calls out limitations and open challenges, which is honest.

Where it gets shaky: the abstract calls this a 'systematic review' and claims to establish a 'foundational taxonomy.' Neither is supported by what is shown. There is no methodology, no search strategy, no selection rationale, no validation of the taxonomy against external benchmarks. That matters because a taxonomy's value depends on whether the sample represents the field's diversity, not just a handful of popular tools. If the full text has a methods section with inclusion/exclusion criteria and some coverage analysis, fine. If the selection was convenience-based, the taxonomy is just an organizing scheme for seven tools and four protocols—still useful, but not foundational. The stress-test note is right to flag this as the least secure assumption.

That said, this is an abstract-only review. An abstract is not a paper, and I am not going to convict the full text on what its abstract left out. The real question is whether the full text backs the rhetoric.

Who is this for? Practitioners choosing an agent framework, and maybe instructors who want a compact comparison. It is not a research contribution in the sense of new theory or data. A serious editor should still send it to peer review: a survey with a taxonomy claim needs a referee to check selection, consistency, and whether the stated limitations match the findings. I would not cite it in my own work within the next year unless it turns out to have actual methodology and I need a convenient reference table. But I would read the full text if you have it.

Referee Report

3 major / 3 minor

Summary. The paper claims to be a systematic review and comparative analysis of seven LLM-based agentic AI frameworks (CrewAI, LangGraph, AutoGen, Semantic Kernel, Agno, Google ADK, MetaGPT) and four agent communication protocols (CNP, A2A, ANP, Agora). It states that it evaluates these systems along architectural principles, communication mechanisms, memory management, safety guardrails, and service-oriented alignment, and that the resulting findings establish a foundational taxonomy for Agentic AI, identify limitations, and propose future research directions for scalability, robustness, and interoperability.

Significance. If the claimed taxonomy is valid and the selected sample is representative, the paper could be a useful consolidating reference for a fragmented and rapidly evolving field. The comparative scope—spanning frameworks and protocols—is timely, and the focus on safety and interoperability addresses practical concerns. However, the abstract alone does not provide evidence for the 'foundational' status of the taxonomy, nor does it show that the evaluation dimensions are complete or that the framework/protocol selection is representative. The paper's significance therefore depends on methodological support that is not visible in the abstract.

major comments (3)

[Abstract (framework/protocol list)] The abstract states that the paper is 'a systematic review' whose findings 'establish a foundational taxonomy,' but it provides no inclusion/exclusion criteria, search strategy, time window, or rationale for selecting these seven frameworks and four protocols. Because the taxonomy's dimensions (architecture, communication, memory, safety, service-orientation) are only meaningful if the sample spans the design space, the claimed foundational status cannot be assessed. Please specify the selection methodology and justify that omitted frameworks (e.g., lightweight agent libraries, non-LLM planners, industrial multi-agent platforms) would not change the taxonomy's categories.
[Abstract ('Our findings not only establish...')] The taxonomy is presented as a result, but there is no indication of how it was derived from the comparative analysis or validated against existing taxonomies. The five evaluation dimensions appear without a derivation or source basis. A load-bearing point for a 'foundational taxonomy' is that the dimensions are complete and not idiosyncratic. Please add a section that maps each dimension to prior literature, explains how the dimensions were chosen, and discusses potential missing dimensions (e.g., human-agent interaction, learning, non-LLM reasoning).
[Abstract ('systematic review and comparative analysis')] The abstract does not specify what kind of evidence the comparative analysis is based on: documentation inspection, code analysis, benchmark experiments, or community adoption metrics. This matters because claims about 'scalability, robustness, and interoperability' can only be supported by a stated evaluation protocol. Clarify the methodology and, if no empirical evaluation was performed, soften the claims to avoid overstatement.

minor comments (3)

[Abstract (terminology)] The acronyms CNP, A2A, ANP, and Agora are not expanded in the abstract; the first use should provide the full names (e.g., Contract Net Protocol, Agent-to-Agent) to help readers.
[Abstract (scope)] The paper is framed as 'systematic review' without indicating the inclusion criteria. If the full text does not contain a protocol, consider replacing 'systematic' with 'comparative survey' to avoid expectations of a formal review methodology.
[Abstract ('foundational taxonomy')] The term 'foundational' is an overclaim for a first comparative survey of seven frameworks. Suggest 'an initial taxonomy' or 'a working taxonomy' unless strong evidence of completeness is provided.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found in abstract-only survey; taxonomy claim is a synthesis, not a derivation from its own outputs.

full rationale

The available material is an abstract only. The paper's claims are descriptive and comparative: it reviews seven frameworks and four protocols, evaluates them along stated dimensions, and proposes a taxonomy and research directions. There is no derivation chain, no fitted parameter renamed as a prediction, and no equation presented that reduces to an input. The 'foundational taxonomy' claim rests on the sufficiency and representativeness of the chosen frameworks and protocols, which is a matter of external validity and sample selection—not circularity. A survey's conclusions are not circular merely because they summarize the surveyed material. No self-citations are visible in the abstract, and there is no evidence that any load-bearing premise is justified solely by the authors' prior work. The caveat that the sample's representativeness is unverified from the abstract is a correctness risk, not a circularity finding. Accordingly, the circularity score is 0.

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

As a review paper, there are no free parameters or invented entities. The load-bearing axioms are the representativeness of the chosen frameworks/protocols and the adequacy of the evaluation criteria. The absence of an explicit methodology in the abstract makes these assumptions unverifiable.

assumptions (3)

domain assumption The seven frameworks and four protocols chosen are representative of the Agentic AI field.
Stated in abstract as 'leading Agentic AI frameworks' and 'in-depth analysis of protocols'.
domain assumption The evaluation dimensions (architectural principles, communication, memory, safety guardrails, SOA alignment) are the appropriate comparative criteria.
Abstract lists these evaluation dimensions without justifying their completeness or relevance.
domain assumption A systematic review methodology exists, even though it is not described in the abstract.
The abstract calls the work a 'systematic review' but provides no details of the selection or analysis process.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Agentic AI Frameworks: Architectures, Protocols, and Design Challenges." pith.science (2026). https://pith.science/paper/6C7WHUCQ

@misc{pith2026250810146,
  author       = {Pith},
  title        = {Pith review of: Agentic AI Frameworks: Architectures, Protocols, and Design Challenges},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6C7WHUCQ}},
  note         = {Machine review of arXiv:2508.10146}
}

read the original abstract

The emergence of Large Language Models (LLMs) has ushered in a transformative paradigm in artificial intelligence, Agentic AI, where intelligent agents exhibit goal-directed autonomy, contextual reasoning, and dynamic multi-agent coordination. This paper provides a systematic review and comparative analysis of leading Agentic AI frameworks, including CrewAI, LangGraph, AutoGen, Semantic Kernel, Agno, Google ADK, and MetaGPT, evaluating their architectural principles, communication mechanisms, memory management, safety guardrails, and alignment with service-oriented computing paradigms. Furthermore, we identify key limitations, emerging trends, and open challenges in the field. To address the issue of agent communication, we conduct an in-depth analysis of protocols such as the Contract Net Protocol (CNP), Agent-to-Agent (A2A), Agent Network Protocol (ANP), and Agora. Our findings not only establish a foundational taxonomy for Agentic AI systems but also propose future research directions to enhance scalability, robustness, and interoperability. This work serves as a comprehensive reference for researchers and practitioners working to advance the next generation of autonomous AI systems.

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

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

Understanding Bugs in Modern Agentic Frameworks: A Study of Symptoms, Root Causes, and Triggering Conditions
cs.SE 2026-04 unverdicted novelty 7.0 of 10

A study of 409 fixed bugs across five agentic frameworks produces taxonomies for symptoms and root causes, finds the model-integration layer is the most bug-prone and least tested, and shows many bug triggers transfer...
Declarative by Design, Assistable Only by Convention: Benchmarking Multi-Agent Frameworks for AI-Assistability
cs.CL 2026-02 reject novelty 6.0 of 10

A controlled benchmark of AI code generation across ten multi-agent frameworks finds that API convention alignment, not declarative design, predicts how well an AI assistant can write correct framework code.
FlexNGIA 2.0: Redesigning the Internet with Agentic AI -- Protocols, Services, and Traffic Engineering Designed, Deployed, and Managed by AI
cs.NI 2025-09 conditional novelty 6.0 of 10

LLM-based agents can autonomously generate working custom transport protocols, congestion-control kernel modules, and resource-allocation weight updates in emulated proof-of-concept experiments.
SoK: How Frontier AI Reshapes System-Level Security Risk Dynamics in Critical Infrastructure
cs.CR 2026-08 conditional novelty 5.0 of 10

A new five-dimension framework describes how frontier AI reshapes critical-infrastructure security through capability, infiltration, propagation, control loss, and response limits.
LLM-Based Agentic Systems for Software Engineering: Challenges and Opportunities
cs.SE 2026-01 unverdicted novelty 2.0 of 10

A survey of LLM-based multi-agent systems across the software development life cycle, plus a research agenda for orchestration, human coordination, cost, and data.

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.