Agoran: An Agentic Open Marketplace for 6G RAN Automation

Ali Maatouk; Andrea Leone; Ilias Chatzistefanidis; Ioannis Pitsiorlas; Leandros Tassiulas; Marios Kountouris; Navid Nikaein; Roberto Morabito

REVIEW 4 major objections 4 minor 4 references

Agoran: An Agentic Open Marketplace for 6G RAN Automation

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper's central claim is that Agoran's three-branch agent marketplace reconciles conflicting slice objectives at run time, and its testbed results show a 37% eMBB throughput gain, a 73% URLLC latency cut, and an 8.3%…

desk verdict Interesting abstract, but the submitted body is a different paper, so the claims are unsupported and the submission is not refereable. read the letter →

arxiv 2508.09159 v2 pith:ITIINRQF submitted 2025-08-05 cs.NI cs.AI

Ilias Chatzistefanidis , Navid Nikaein , Andrea Leone , Ali Maatouk , Leandros Tassiulas , Roberto Morabito , Ioannis Pitsiorlas , Marios Kountouris This is my paper

classification cs.NIcs.AI

keywords agenticmarketplace6GRANautomationnetworkslicingmulti-objectiveoptimizationLLM-basednegotiationOpentrustscoringserviceandresourcebroker

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 tries to establish that the business layer of a mobile network—the service owners with conflicting goals—can be brought into the operating loop of radio access network (RAN) slicing, instead of being locked out by rigid policy-bound controllers. It proposes Agoran, an agentic marketplace with three autonomous branches: a Legislative branch that answers compliance queries with retrieval-augmented language models, an Executive branch that maintains real-time situational awareness through a watcher-updated vector database, and a Judicial branch that scores trust and arbitrates suspicious behavior. Stakeholder-side negotiation agents and a mediator use a multi-objective optimizer to produce Pareto-optimal offers and reach a single consensus intent, which is then deployed to Open and AI RAN controllers. On a private 5G testbed with realistic vehicle-mobility traces, the authors report a 37% increase in eMBB slice throughput, a 73% reduction in URLLC slice latency, and an 8.3% end-to-end saving in physical resource block usage versus a static baseline. A 1B-parameter Llama model fine-tuned for five minutes on 100 GPT-4 dialogues is reported to recover about 80% of GPT-4.1's decision quality while running within 6 GiB of memory and converging in 1.3 seconds.

What carries the argument

The load-bearing mechanism is the Agoran Service and Resource Broker (SRB), a three-branch agentic marketplace. The Legislative branch uses retrieval-augmented large language models to answer compliance queries; the Executive branch keeps a watcher-updated vector database as its situational-awareness store; and the Judicial branch scores each agent message with a rule-based Trust Score, with arbitrating LLMs detecting malicious behavior and applying real-time incentives to restore trust. The economic core is a multi-objective optimizer that generates Pareto-optimal offers for the stakeholder-side Negotiation Agents and the SRB-side Mediator Agent, whose single-round consensus intent is the artifact actually deployed to Open and AI RAN controllers.

What would settle it

Re-run the same 5G testbed with deliberately conflicting slice objectives—for example, a URLLC latency target that is infeasible under current radio conditions alongside an eMBB throughput demand that exhausts nearly all physical resource blocks—and count how many negotiation rounds end without an accepted consensus intent. A substantial rejection or renegotiation rate would show the single-round mechanism is not robust.

Watch

Extended reading notes

Core claim

The central claim is that conflicting service-owner objectives can be reconciled at run time through an agentic marketplace rather than through static, policy-bound slice control. In Agoran, the negotiation loop is closed by a single-round consensus intent: the mediator and the stakeholder-side agents agree on one feasible, Pareto-optimal offer, and that intent is pushed directly to Open and AI RAN controllers. The paper's evidence is the private 5G testbed evaluation—eMBB throughput up 37%, URLLC latency down 73%, and end-to-end PRB usage down 8.3% against a static baseline—together with the demonstration that a 1B-parameter Llama model can reproduce about 80% of the larger model's decision quality in 1.3 seconds.

Load-bearing premise

The load-bearing premise is that the stakeholder-side negotiation agents and the mediator always reach a single consensus intent in one round, and that Open and AI RAN controllers accept and apply that intent directly; if that handshake fails or requires revision, the marketplace loop and the reported gains break down.

Editorial extensions

If this is right

Run-time reconciliation of slice owners becomes possible: the negotiation loop is designed to be fast enough for mobility-driven demand changes, with convergence reported at 1.3 seconds.
Small, cheaply trained models can substitute for frontier models in RAN arbitration, since the 1B-parameter model recovers roughly 80% of the larger model's decision quality.
The marketplace's consensus intents are deployable to Open and AI RAN controllers, so the approach is a standards-aligned evolution path rather than a proprietary overlay.
The judicial trust loop gives operators a concrete way to detect malicious agents and steer them back to cooperative behavior with incentives, which is a prerequisite for opening the RAN to outside stakeholders.
The end-to-end 8.3% PRB saving means shared spectrum and radio resources can be used more efficiently while service owners negotiate, not just at the scheduling layer.

Reading between the lines

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

If the single-round consensus assumption is stress-tested with intentionally incompatible slice demands, the mediator will sometimes need a second round or must drop an offer; the paper does not report those failure rates.
The same three-branch marketplace pattern could apply to other shared infrastructures—cloud schedulers, spectrum brokers, or energy grids—where multiple owners bid for hard-latency resources.
The 80% quality recovery at 1B parameters suggests frontier-model distillation may lower the operational cost of RAN automation, but the paper does not evaluate it under non-stationary traffic or adversarial agents.
Since trust scoring is rule-based, the judicial branch's incentives are only as good as those rules; a learned trust model is a natural extension the paper leaves unexplored.

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

The punchline is that the abstract describes an agentic marketplace for 6G RAN automation with real deployment numbers, but the full text is a completely different paper about dense layers in time-series models. So as submitted, the paper is not reviewable: every performance claim in the abstract—37% throughput gain, 73% latency reduction, 8.3% PRB saving, 80% GPT-4.1 decision recovery by a 1B Llama—appears nowhere in the body. There is no architecture, no experimental protocol, no baseline definition, no error bars, no code. I can't verify a single number.

What is genuinely interesting, if the abstract is accurate, is the architectural pattern: three AI branches (Legislative, Executive, Judicial) with RAG-LLMs, a rule-based trust score, and a mediator that negotiates Pareto-optimal offers between stakeholders. That combination is new to me in RAN control. The small-LLM distillation result—fine-tuning a 1B model for five minutes on 100 GPT-4 dialogues to recover 80% of decision quality—is the kind of concrete, falsifiable claim that would matter if backed by a real testbed.

The soft spots are large. First, the full-text mismatch is fatal for peer review. A serious editor would desk reject this submission on the spot; the body has nothing to do with the abstract. Second, even taking the abstract on its own, the 'consensus intent in a single round' is an unmotivated assumption. Multi-objective negotiation with complete utility revelation in one round is a strong condition, and the abstract doesn't say what happens if the controllers reject the intent. Third, there are no references in the abstract, so novelty cannot be checked. Self-citation is not the issue; there is simply nothing to benchmark against.

For whom is this useful? For a reader working on LLM-driven RAN control, the abstract is a one-paragraph teaser that might justify a closer look at a later, complete version. But this submission, in its current form, deserves no referee time. The right move is to return it as incomplete and ask for the actual manuscript.

Referee Report

4 major / 4 minor

Summary. The manuscript submitted under arXiv:2508.09159 (cs.NI) presents Agoran, an agentic marketplace for 6G RAN automation, in its abstract. The abstract describes a three-branch AI architecture (Legislative, Executive, Judicial), stakeholder-side negotiation agents, a mediator agent, a multi-objective optimizer, and deployment on a private 5G testbed, with claimed gains of 37% in eMBB throughput, 73% in URLLC latency reduction, and 8.3% in PRB usage, plus a fine-tuned 1B Llama model recovering 80% of GPT-4.1's decision quality. However, the full text supplied with the submission is a different paper, 'JustDense: Just using Dense instead of Sequence Mixer for Time Series analysis' (arXiv:2508.09153v1), which contains no description of Agoran, its architecture, its experimental protocol, its baseline, or any of the claimed results. The central claims of the abstract therefore have no supporting methodology or evidence in the submitted body.

Significance. If the Agoran results were fully documented, the work could be of interest to the RAN automation community as a concrete proposal for stakeholder-driven slice negotiation with LLM-based compliance checking and trust management. The claimed end-to-end gains on a private 5G testbed and the resource-efficient LLM compression result would be useful if reproducible. However, in the current submission, none of this is verifiable: the full text is an unrelated time-series study, and no architecture, algorithms, experimental setup, baseline details, error bars, or code are provided for Agoran. The paper therefore cannot currently make a contribution to the literature, and the significance assessment is conditional on a complete resubmission that matches the abstract.

major comments (4)

[Full Text (all sections)] The submitted full text is a completely different manuscript, 'JustDense: Just using Dense instead of Sequence Mixer for Time Series analysis' (arXiv:2508.09153v1), and contains no mention of Agoran, the Service and Resource Broker, negotiation agents, Open/AI RAN controllers, 5G testbeds, vehicle mobility traces, or any of the claimed experimental results. Consequently, none of the central claims in the abstract — 37% eMBB throughput increase, 73% URLLC latency reduction, 8.3% PRB saving, and 80% GPT-4.1 decision-quality recovery — have any supporting methodology or evidence in the body. This is a load-bearing defect that prevents verification of the paper's central claims.
[Abstract (Agoran architecture)] The abstract asserts that Negotiation Agents and the Mediator Agent 'reach a consensus intent in a single round' after multi-objective optimization, but the submission provides no mechanism, formal description, or experimental evidence for this convergence. In particular, there is no discussion of the information stakeholders must reveal, the conditions under which a Pareto-optimal offer is acceptable to all parties, or what happens if the Open and AI RAN controllers reject the negotiated intent. Since the end-to-end loop depends on this single-round consensus, this unstated assumption is load-bearing and cannot be assessed from the submitted text.
[Abstract (experimental claims)] The quantitative results (37%, 73%, 8.3%, and the LLM's 80% recovery) are presented without any experimental protocol: there is no description of the private 5G testbed, the baseline configuration, the traffic models, the number of runs, the variance or confidence intervals, or the exact metrics used. The LLM claim additionally lacks details of the fine-tuning procedure, the evaluation set, the notion of 'decision quality', and the comparison to GPT-4.1. These omissions make the reported numbers neither reproducible nor statistically assessable.
[Abstract (LLM compression)] The claim that a 1B-parameter Llama model fine-tuned for five minutes on 100 GPT-4 dialogues recovers approximately 80% of GPT-4.1's decision quality is unsupported by any methodology. No information is given about the dialogue generation process, the fine-tuning objective, the hardware used for the 1.3-second convergence figure, the 6 GiB memory measurement, or the evaluation benchmark. As stated, this is an isolated assertion with no way to verify or compare it.

minor comments (4)

[Abstract] The phrase 'standards-aligned' is used without naming any standard or describing the alignment procedure; the reader cannot determine which 3GPP or O-RAN specifications are being referenced.
[Abstract] The terms 'Legislative branch', 'Executive branch', and 'Judicial branch' are introduced without definitions, examples, or pseudocode, making the architecture hard to follow even before the missing full-text support.
[Abstract] The live demo URL is mentioned but no demonstration content, duration, or behavior is described; a video link cannot substitute for an experimental section.
[Full Text (JustDense)] The submitted body's title, authors, and content are inconsistent with the arXiv metadata and abstract of the Agoran paper, which is a presentation-level issue that also reflects the lack of a coherent manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the abstract/full-text mismatch is a verifiability concern, not a circular derivation.

full rationale

The submitted abstract describes the Agoran marketplace, while the full text is an unrelated time-series paper (JustDense). However, the circularity analysis requires exhibiting a specific reduction where a claimed prediction or derivation is equivalent to its inputs by construction, or where a self-citation carries the load. No such reduction appears in the abstract: the reported eMBB throughput, URLLC latency, PRB savings, and LLM decision-quality recovery are stated as empirical results compared to a static baseline or to GPT-4.1 decisions, respectively. The LLM fine-tuning claim is a distillation-style evaluation against the teacher model, which is a direct comparison rather than a constructed equivalence. The 'single-round consensus intent' is an assumption about negotiation behavior, not a definitional identity. The mismatch between the abstract and the full text is a serious verifiability and correctness issue, because the body provides no experimental protocol or architecture for Agoran, but it is not circularity under the criteria required here. Therefore, the circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 3 invented entities

Because only the abstract is available, all free parameters, axioms, and entities are inferred from the text. The actual values and assumptions are likely more numerous in the full paper.

free parameters (4)

Trust Score thresholds
The rule-based Trust Score likely depends on thresholds for classifying agent messages as malicious; values are not disclosed in the abstract.
Incentive parameters
Real-time incentives to restore trust have adjustment parameters; not specified in the abstract.
Multi-objective optimizer weights
Pareto-optimal offers require weighting of throughput, latency, and resource efficiency; weights not disclosed.
LLM fine-tuning hyperparameters = 5 minutes, 100 dialogues
The abstract states fine-tuning time and data size, but not learning rate, batch size, or other hyperparameters.

assumptions (5)

domain assumption Open and AI RAN controllers expose sufficient control interfaces to deploy negotiated intents
The deployment step assumes these interfaces exist and are reliable, based on the abstract's mention of Open and AI RAN controllers.
domain assumption LLM-based Legislative branch answers compliance queries accurately enough for operational decisions
The abstract relies on retrieval-augmented LLMs for compliance, but does not quantify error rates.
domain assumption Watcher-updated vector database reflects real-time network state accurately
The Executive branch's situational awareness depends on freshness and correctness of the vector database.
domain assumption Rule-based Trust Score and arbitration LLMs can reliably distinguish malicious behavior
Security and trust restoration depend on this detection capability, which is asserted without evidence.
ad hoc to paper Pareto-optimal offers are negotiable to consensus in one round
Single-round consensus is a design guarantee that is central to the marketplace but not derived from any underlying principle in the abstract.

invented entities (3)

Agoran Service and Resource Broker (SRB)
purpose: Acts as an agentic marketplace coordinating stakeholders, negotiating slices, and enforcing compliance and trust.
The SRB is the proposed system. The abstract reports testbed results, but these are internal to the paper and no external benchmark is given.
Legislative, Executive, and Judicial AI branches
purpose: Divide authority for compliance, situational awareness, and trust enforcement.
These are the system's components, not observed independently.
Negotiation Agents and Mediator Agent
purpose: Represent stakeholders and broker consensus intents.
Software agents central to the mechanism, no external validation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Agoran: An Agentic Open Marketplace for 6G RAN Automation." pith.science (2026). https://pith.science/paper/ITIINRQF

@misc{pith2026250809159,
  author       = {Pith},
  title        = {Pith review of: Agoran: An Agentic Open Marketplace for 6G RAN Automation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ITIINRQF}},
  note         = {Machine review of arXiv:2508.09159}
}

read the original abstract

Next-generation mobile networks must reconcile the often-conflicting goals of multiple service owners. However, today's network slice controllers remain rigid, policy-bound, and unaware of the business context. We introduce Agoran Service and Resource Broker (SRB), an agentic marketplace that brings stakeholders directly into the operational loop. Inspired by the ancient Greek agora, Agoran distributes authority across three autonomous AI branches: a Legislative branch that answers compliance queries using retrieval-augmented Large Language Models (LLMs); an Executive branch that maintains real-time situational awareness through a watcher-updated vector database; and a Judicial branch that evaluates each agent message with a rule-based Trust Score, while arbitrating LLMs detect malicious behavior and apply real-time incentives to restore trust. Stakeholder-side Negotiation Agents and the SRB-side Mediator Agent negotiate feasible, Pareto-optimal offers produced by a multi-objective optimizer, reaching a consensus intent in a single round, which is then deployed to Open and AI RAN controllers. Deployed on a private 5G testbed and evaluated with realistic traces of vehicle mobility, Agoran achieved significant gains: (i) a 37% increase in throughput of eMBB slices, (ii) a 73% reduction in latency of URLLC slices, and concurrently (iii) an end-to-end 8.3% saving in PRB usage compared to a static baseline. An 1B-parameter Llama model, fine-tuned for five minutes on 100 GPT-4 dialogues, recovers approximately 80% of GPT-4.1's decision quality, while operating within 6 GiB of memory and converging in only 1.3 seconds. These results establish Agoran as a concrete, standards-aligned path toward ultra-flexible, stakeholder-centric 6G networks. A live demo is presented https://www.youtube.com/watch?v=h7vEyMu2f5w\&ab_channel=BubbleRAN.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

4 extracted references · 1 canonical work pages

[2]

InInternational Conference on Learning Representations

A Time Series is Worth 64 Words: Long-term Fore- casting with Transformers. InInternational Conference on Learning Representations. NVIDIA Corporation. 2025.cuBLAS Library Programmer’s Guide. Santa Clara, CA, version 12.9 edition. Release 12.9, accessed 27 Jul 2025. Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z...

work page 2025
[2019]

InProceed- ings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2828–2837

Robust anomaly detection for multivariate time series through stochastic recurrent neural network. InProceed- ings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2828–2837. Thomas, A.; Gu, A.; Dao, T.; Rudra, A.; and R ´e, C. 2018. Learning compressed transforms with low displacement rank.Advances in neural informati...

arXiv 2018
[2023]

Capistr´an, C.; Constandse, C.; and Ramos-Francia, M

Accurate medium-range global weather forecasting with 3D neural networks.Nature, 619(7970): 533–538. Capistr´an, C.; Constandse, C.; and Ramos-Francia, M. 2010. Multi-horizon inflation forecasts using disaggregated data. Economic Modelling, 27(3): 666–677. Chen, S.-A.; Li, C.-L.; Yoder, N.; Arik, S. O.; and Pfister, T. 2023. Tsmixer: An all-mlp architectu...

arXiv 2010
[2024]

Wang, Z.; Kong, F.; Feng, S.; Wang, M.; Yang, X.; Zhao, H.; Wang, D.; and Zhang, Y

Deep Time Series Models: A Comprehensive Survey and Benchmark.CoRR. Wang, Z.; Kong, F.; Feng, S.; Wang, M.; Yang, X.; Zhao, H.; Wang, D.; and Zhang, Y . 2025. Is mamba effective for time series forecasting?Neurocomputing, 619: 129178. Wu, H.; Xu, J.; Wang, J.; and Long, M. 2021. Auto- former: Decomposition Transformers with Auto-Correlation for Long-Term ...

arXiv 2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [2]

InInternational Conference on Learning Representations

A Time Series is Worth 64 Words: Long-term Fore- casting with Transformers. InInternational Conference on Learning Representations. NVIDIA Corporation. 2025.cuBLAS Library Programmer’s Guide. Santa Clara, CA, version 12.9 edition. Release 12.9, accessed 27 Jul 2025. Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z...

work page 2025

[2] [2019]

InProceed- ings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2828–2837

Robust anomaly detection for multivariate time series through stochastic recurrent neural network. InProceed- ings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2828–2837. Thomas, A.; Gu, A.; Dao, T.; Rudra, A.; and R ´e, C. 2018. Learning compressed transforms with low displacement rank.Advances in neural informati...

arXiv 2018

[3] [2023]

Capistr´an, C.; Constandse, C.; and Ramos-Francia, M

Accurate medium-range global weather forecasting with 3D neural networks.Nature, 619(7970): 533–538. Capistr´an, C.; Constandse, C.; and Ramos-Francia, M. 2010. Multi-horizon inflation forecasts using disaggregated data. Economic Modelling, 27(3): 666–677. Chen, S.-A.; Li, C.-L.; Yoder, N.; Arik, S. O.; and Pfister, T. 2023. Tsmixer: An all-mlp architectu...

arXiv 2010

[4] [2024]

Wang, Z.; Kong, F.; Feng, S.; Wang, M.; Yang, X.; Zhao, H.; Wang, D.; and Zhang, Y

Deep Time Series Models: A Comprehensive Survey and Benchmark.CoRR. Wang, Z.; Kong, F.; Feng, S.; Wang, M.; Yang, X.; Zhao, H.; Wang, D.; and Zhang, Y . 2025. Is mamba effective for time series forecasting?Neurocomputing, 619: 129178. Wu, H.; Xu, J.; Wang, J.; and Long, M. 2021. Auto- former: Decomposition Transformers with Auto-Correlation for Long-Term ...

arXiv 2025