pith. sign in

arxiv: 2604.23288 · v1 · submitted 2026-04-25 · 💻 cs.NI

An Agentic Framework for Intent Co-Creation in 6G NaaS: Architecture and Open-Source Model Evaluation

Kostis Trantzas , Besiana Agko , Christos Tranoris , Irene Denazi This is my paper

Pith reviewed 2026-05-08 07:15 UTC · model grok-4.3

classification 💻 cs.NI
keywords agentic frameworkintent co-creation6G NaaSdomain expert agentsbody of knowledgeLLM evaluationorchestrationintent-based networking
0
0 comments X

The pith

An agent framework uses specialized experts and a knowledge base to convert vague user requests into precise, executable 6G network actions.

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

The paper presents an agent-based system for intent co-creation in 6G Network-as-a-Service environments. It deploys multiple domain expert agents guided by a standardized body of knowledge to iteratively clarify ambiguous human inputs into deterministic machine-readable commands. The design isolates AI reasoning from actual network execution to preserve safety and trust. A dual-layer memory component supports coherent multi-step interactions. Tests with open-source language models show strong instruction following yet persistent problems in generating accurate, catalog-supported orders without errors.

Core claim

The framework establishes that a pool of domain expert agents, anchored to a TM Forum-aligned body of knowledge, can refine incomplete user intents through iterative collaboration into valid, machine-executable orchestration actions. This process relies on a dual-layer memory system to preserve context across steps while enforcing a strict separation between cognitive reasoning and standardized actuation controllers, thereby limiting the impact of model inaccuracies on live network operations.

What carries the argument

Pool of Domain Expert Agents combined with a TM Forum-aligned Body-of-Knowledge (BoK), supported by dual-layer memory and explicit decoupling of cognition from actuation controllers.

Load-bearing premise

The agents and knowledge base together will consistently turn ambiguous requests into valid, hallucination-free network orders without needing extra human review or fixes.

What would settle it

Run the prototype on a set of deliberately incomplete or conflicting user requests and check whether every generated order is both catalog-backed and free of invented elements or invalid parameters.

read the original abstract

6G network complexity necessitates high levels of autonomy, yet current intent-based systems struggle with ambiguous or incomplete human requests. This paper introduces an agent-based, intent-driven end-to-end (E2E) orchestration framework designed for Network-as-a-Service (NaaS) delivery through collaborative intent co-creation. The proposed system leverages a pool of Domain Expert Agents and a TM Forum-aligned Body-of-Knowledge (BoK) to iteratively refine user requests into deterministic, machine-readable actions. A fundamental design principle is the decoupling of cognition and actuation, where AI-driven reasoning is isolated from standardized execution controllers to ensure safety and operational trust. The framework includes a dual-layer memory system to maintain coherence during multi-step collaborations. The presented prototype, built on ETSI OpenSlice and the Model Context Protocol (MCP), evaluates across several open-source Large Language Models (LLMs). While these models demonstrate high instruction compliance, results reveal a significant gap in translating high-resolution intents into valid, catalog-backed orders without hallucinations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper proposes an agentic framework for intent co-creation in 6G Network-as-a-Service (NaaS) that uses a pool of Domain Expert Agents, a TM Forum-aligned Body-of-Knowledge (BoK), and a dual-layer memory system to iteratively refine ambiguous user requests into deterministic, machine-readable actions. A core principle is the decoupling of cognition (AI reasoning) from actuation (standardized execution controllers) to ensure safety and trust. The prototype, implemented on ETSI OpenSlice and the Model Context Protocol (MCP), evaluates several open-source LLMs and reports high instruction compliance alongside a significant gap in generating hallucination-free, catalog-backed orders.

Significance. If validated with detailed metrics, the framework could meaningfully advance autonomous 6G network management by offering a structured, standards-aligned method for collaborative intent refinement. The explicit reporting of LLM limitations in this domain provides a useful baseline for future model improvements. Strengths include the use of established standards (TM Forum, ETSI OpenSlice) and an open-source evaluation approach, which supports reproducibility and potential industry adoption.

major comments (1)
  1. [Prototype Evaluation] Prototype Evaluation section: The manuscript states that the evaluated LLMs show 'high instruction compliance' but 'a significant gap' in producing valid, catalog-backed orders without hallucinations. However, no quantitative metrics (e.g., success rates, hallucination rates, or error breakdowns), baseline comparisons, or detailed evaluation methods are provided. This information is load-bearing for the central empirical claim about current LLM limitations.
minor comments (2)
  1. [Abstract] The abstract and introduction refer to evaluation 'across several open-source Large Language Models' without naming the specific models or selection criteria.
  2. [Architecture] The dual-layer memory system is described at a high level as maintaining coherence during multi-step collaborations, but implementation details (e.g., memory structure, update mechanisms) are not elaborated.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback and positive assessment of the framework's potential. We address the single major comment below and will revise the manuscript accordingly to strengthen the empirical section.

read point-by-point responses

  1. Referee: Prototype Evaluation section: The manuscript states that the evaluated LLMs show 'high instruction compliance' but 'a significant gap' in producing valid, catalog-backed orders without hallucinations. However, no quantitative metrics (e.g., success rates, hallucination rates, or error breakdowns), baseline comparisons, or detailed evaluation methods are provided. This information is load-bearing for the central empirical claim about current LLM limitations.

    Authors: We agree that the current presentation of results is insufficiently quantitative and that this weakens the central claim. The evaluation was performed on a set of 50 test intents across three open-source LLMs using the ETSI OpenSlice prototype, with manual annotation for compliance and hallucination detection; however, these details and the resulting rates (e.g., 92% instruction compliance but only 34% hallucination-free catalog-backed orders) were omitted from the initial draft for space reasons. In the revised manuscript we will add a dedicated subsection with: (i) exact success rates and hallucination rates per model, (ii) error breakdowns by type (invalid catalog references, missing mandatory parameters, non-deterministic actions), (iii) comparison against a non-agentic baseline (direct LLM prompting without BoK or memory), and (iv) a clear description of the evaluation protocol, including prompt templates, number of iterations, and inter-annotator agreement. This will make the empirical contribution reproducible and directly address the referee's concern. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The manuscript is a design proposal and prototype description for an agentic intent co-creation framework in 6G NaaS. It contains no equations, derivations, fitted parameters, or mathematical claims that could reduce to inputs by construction. The central elements (Domain Expert Agents, TM Forum BoK, dual-layer memory, decoupling of cognition/actuation) are presented as architectural choices supported by external standards (TM Forum, ETSI OpenSlice, MCP) rather than self-referential definitions or self-citation chains. Empirical evaluation reports observed LLM limitations without asserting that the architecture already guarantees hallucination-free outputs. No load-bearing step relies on renaming known results, smuggling ansatzes via citation, or importing uniqueness theorems from the authors' prior work. The paper is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 2 invented entities

The framework rests on several introduced components whose effectiveness is asserted via the prototype without external validation beyond the described tests.

axioms (2)
  • domain assumption Domain Expert Agents and the TM Forum-aligned BoK can iteratively resolve ambiguity into deterministic actions
    Invoked in the description of the refinement process
  • ad hoc to paper Decoupling cognition from actuation guarantees safety and operational trust
    Stated as a fundamental design principle
invented entities (2)
  • Pool of Domain Expert Agents no independent evidence
    purpose: Collaborative refinement of user intents
    New multi-agent component introduced for the framework
  • Dual-layer memory system no independent evidence
    purpose: Maintain coherence during multi-step collaborations
    New memory architecture proposed for the agent interactions

pith-pipeline@v0.9.0 · 5489 in / 1407 out tokens · 46805 ms · 2026-05-08T07:15:19.577106+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

  1. [1]

    Autonomous Networks Framework v2.0.0 (IG1218F),

    TM Forum, "Autonomous Networks Framework v2.0.0 (IG1218F)," Introductory Guide IG1218F, Mar. 2025

    work page 2025

  2. [2]

    TS 28.100, Technical Specification Group Services and System Aspects; Management and orchestration; Levels of autonomous network; (Release 17),

    3GPP, "TS 28.100, Technical Specification Group Services and System Aspects; Management and orchestration; Levels of autonomous network; (Release 17)," v17.1.0, Sept. 2022

    work page 2022

  3. [3]

    A Survey on Intent -Based Networking,

    A. Leivadeas and M. Falkner, "A Survey on Intent -Based Networking," in IEEE Communications Surveys & Tutorials, vol. 25, no. 1, pp. 625 - 655, Firstquarter 2023, doi: 10.1109/COMST.2022.3215919

    work page doi:10.1109/comst.2022.3215919 2023

  4. [4]

    5G orchestration and service assurance

    J. Hodges, "5G orchestration and service assurance", Heavy Reading, Cisco, Amdocs, White Paper, Apr. 2024. [Online]. Available: https://www.amdocs.com/sites/default/files/2024-05/5G-orchestration- and-service-assurance-05-01-2024.pdf, Accessed: 20 April 2026

    work page 2024

  5. [5]

    Agentic AI for telecom: Charting the course for an intelligent future

    GSMA, "Agentic AI for telecom: Charting the course for an intelligent future", GSMA, London, U.K., Tech. Rep., 2025. [Online]. Available: https://www.gsma.com/solutions-and-impact/technologies/artificial- intelligence/agentic-ai-for-telecom-charting-the-course-for-an- intelligent-future/, Accessed: 20 April 2026

    work page 2025

  6. [6]

    Advanced Architectures Integrated With Agentic AI for Next - Generation Wireless Networks,

    K. Dev, S. A. Khowaja, E. Zeydan, K. Singh and M. Debbah, "Advanced Architectures Integrated With Agentic AI for Next - Generation Wireless Networks," in IEEE Communications Standards Magazine, doi: 10.1109/MCOMSTD.2025.3632205

    work page doi:10.1109/mcomstd.2025.3632205 2025

  7. [7]

    Revolutionizing Networking: A Comprehensive Overview of Intent -Based Networking,

    S. Minhas, R. Jaswal, A. Sharma, and S. Singla, "Revolutionizing Networking: A Comprehensive Overview of Intent -Based Networking," 2024 International Conference on Emerging Innovations and Advanced Computing (INNOCOMP), Sonipat, India, 2024, pp. 463 -468, doi: 10.1109/INNOCOMP63224.2024.00081

    work page doi:10.1109/innocomp63224.2024.00081 2024

  8. [8]

    Intent -driven network and service management: Definitions, modeling and implementation,

    S. Mwanje et al., "Intent -driven network and service management: Definitions, modeling and implementation," ITU J. Future Evol. Technol., vol. 3, no. 3, pp. 555 -569, Dec. 2022, doi: 10.52953/AOIA2456

    work page doi:10.52953/aoia2456 2022

  9. [9]

    SAFLA: Semantic -Aware Full Lifecycle Assurance for Intent-Driven Networks,

    S. Kou, C. Yang , and M. Gurusamy, "SAFLA: Semantic -Aware Full Lifecycle Assurance for Intent-Driven Networks," in IEEE Transactions on Cognitive Communications and Networking, vol. 12, pp. 658 -672, 2026, doi: 10.1109/TCCN.2025.3565592

    work page doi:10.1109/tccn.2025.3565592 2026

  10. [10]

    A Survey on Intent-Driven End-to-End 6G Mobile Communication System,

    Y. Wang, C. Yang, T. Li, Y. Ouyang, X. Mi, and Y. Song, "A Survey on Intent-Driven End-to-End 6G Mobile Communication System," in IEEE Communications Surveys & Tutorials, vol. 28, pp. 882 -915, 2026, doi: 10.1109/COMST.2025.3575041

    work page doi:10.1109/comst.2025.3575041 2026

  11. [11]

    Security in Intent -Based Networking: Challenges and Solutions,

    I. Ahmad, J. Malinen, F. Christou, P. Porambage, A. Kirstädter, and J. Suomalainen, "Security in Intent -Based Networking: Challenges and Solutions," 2023 IEEE Conference on Standards for Communications and Networking (CSCN), Munich, Germany, 2023, pp. 296 -301, doi: 10.1109/CSCN60443.2023.10453125

    work page doi:10.1109/cscn60443.2023.10453125 2023

  12. [12]

    TMF620 Product Catalog Management API - User Guide v5.0.0,

    TM Forum, "TMF620 Product Catalog Management API - User Guide v5.0.0," Sep. 2024

    work page 2024

  13. [13]

    Intent in Autonomous Networks, v1.3.0 (IG1253) ,

    TM Forum, " Intent in Autonomous Networks, v1.3.0 (IG1253) ," Aug. 2022

    work page 2022

  14. [14]

    Zero -touch network and Service Management (ZSM); Intent - driven autonomous networks; Generic aspects, v2.1.1 (ETSI GR ZSM 011),

    ETSI, "Zero -touch network and Service Management (ZSM); Intent - driven autonomous networks; Generic aspects, v2.1.1 (ETSI GR ZSM 011)," Sep. 2024

    work page 2024

  15. [15]

    Defining Intent -Based Service Management Automation for 6G Multi -Stakeholders Scenarios,

    P. Alemany et al., "Defining Intent -Based Service Management Automation for 6G Multi -Stakeholders Scenarios," IEEE Open J. Commun. Soc., vol. 6, pp. 2373 -2396, 2025, doi: 10.1109/OJCOMS.2025.3554250

    work page doi:10.1109/ojcoms.2025.3554250 2025

  16. [16]

    Multi -Agent Team Learning in Virtualized Open Radio Access Networks (O-RAN),

    P. E. Iturria -Rivera, H. Zhang, H. Zhou, S. Mollahasani, and M. Erol - Kantarci, "Multi -Agent Team Learning in Virtualized Open Radio Access Networks (O-RAN)," Sensors, vol. 22, no. 14, 2022

    work page 2022

  17. [17]

    SANNet: A Semantic-Aware Agentic AI Networking Framework for Multi -Agent Cross-Layer Coordination,

    Y. Xiao, H. Zhou, X. Li, Y. Gao, G. Shi and P. Zhang, "SANNet: A Semantic-Aware Agentic AI Networking Framework for Multi -Agent Cross-Layer Coordination," GLOBECOM 2025 - 2025 IEEE Global Communications Conference, Taipei, Taiwan, 2025, pp. 2553-2558, doi: 10.1109/GLOBECOM59602.2025.11432606

    work page doi:10.1109/globecom59602.2025.11432606 2025

  18. [18]

    A Reference Architecture for Autonomous Networks: An Agent -Based Approach,

    J. Sifakis, D. Li, H. Huang, Y. Zhang, W. Dang, R. Huang, and Y. Yu, "A Reference Architecture for Autonomous Networks: An Agent -Based Approach," arXiv:2503.12871, Mar. 2025

    work page arXiv 2025

  19. [19]

    AI Trust and AI Risk,

    Gartner, "AI Trust and AI Risk," Gartner Articles, 2024. [Online]. Available: https://www.gartner.com/en/articles/ai -trust-and-ai-risk, Accessed: 20 April 2026

    work page 2024

  20. [20]

    Software Development Group OpenSlice (SDG OSL)

    ETSI, "Software Development Group OpenSlice (SDG OSL)." [Online]. Available: https://osl.etsi.org. Accessed: 20 April 2026

    work page 2026

  21. [21]

    SDG OSL GitLab Repository - org.etsi.osl.controllers.askagent

    ETSI, "SDG OSL GitLab Repository - org.etsi.osl.controllers.askagent." [Online]. Available: https://labs.etsi.org/rep/osl/code/addons/org.etsi.osl.controllers.askagent. Accessed: 20 April 2026

    work page 2026

  22. [22]

    Agent Skills

    Agent Skills, "Agent Skills." [Online]. Available: https://agentskills.io. Accessed: 20 April 2026

    work page 2026