pith. sign in

arxiv: 2604.13384 · v1 · submitted 2026-04-15 · 💻 cs.NI

Agentic Open RAN: A Deterministic and Auditable Framework for Intent-Driven Radio Control

Hengxu Li , Dongkuan Xu , Mingzhe Chen , Yuchen Liu This is my paper

Pith reviewed 2026-05-10 12:51 UTC · model grok-4.3

classification 💻 cs.NI
keywords Open RANAgentic controlIntent-driven networkingDeterministic executionRIC coordinationPolicy compilationLLM agentsRadio resource management
0
0 comments X

The pith

A1gent decouples natural-language intent reasoning from deterministic near-real-time execution to deliver auditable agentic control in Open RAN.

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

The paper presents A1gent as a control stack that lets network operators state goals in everyday language while autonomous agents translate and enforce them in the radio access network. It achieves this by placing the language-model reasoning in a non-real-time component that produces typed policies, then hands them to near-real-time agents that execute fixed, deterministic loops on mobility, load, and energy planes. Guardrails, priority-based conflict resolution, and a memory-based tuner that avoids retraining keep the system predictable and verifiable. A sympathetic reader would care because this approach could turn complex radio networks into systems that follow high-level instructions reliably and can be audited after the fact.

Core claim

A1gent advances Open RAN by integrating intent-driven planning with deterministic near-RT execution. A non-RT agentic rApp compiles operator goals into typed A1 policy instances. Three task-oriented near-RT agentic xApps enforce them through a deterministic loop with plane-scoped actuation using E2 for mobility and load steering and O1 for energy orchestration. Auditable coordination is supported by encoded guardrails and a fixed-priority action merger, while a training-free adaptive policy tuner refines parameters using KPI memory.

What carries the argument

The reasoning-execution split between non-RT agentic rApp for intent compilation into typed policies and near-RT agentic xApps for deterministic enforcement with guardrails and priority merging.

If this is right

  • Operators can express high-level goals in natural language that translate directly into enforceable typed policies.
  • Real-time actions remain predictable because enforcement uses fixed deterministic loops and memory-based tuning without retraining.
  • Action conflicts are resolved through fixed-priority merging and encoded guardrails to maintain safe operation.
  • Coordination between non-RT and near-RT RIC tiers produces auditable records of intent translation and execution.
  • Open RAN systems gain a path toward self-governing behavior that operators can verify and reproduce.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same split between flexible planning and strict execution could apply to other real-time control domains such as cloud resource management or industrial automation.
  • Ambiguous natural-language intents would likely require explicit human confirmation loops before policy compilation to prevent downstream errors.
  • Standardized A1 policy schemas could become a practical requirement for future Open RAN releases if this pattern is adopted.
  • Simulation-based validation of the non-RT compilation step against known edge-case intents would be a direct way to test the weakest assumption before live trials.

Load-bearing premise

The non-RT agentic rApp can reliably compile natural-language intents into typed A1 policies that the near-RT xApps can enforce without introducing unacceptable latency or safety violations in live networks.

What would settle it

A deployment test in which compiled policies produce measurable latency violations, safety breaches, or failure to meet the stated intent would show that the split does not deliver reliable control.

Figures

Figures reproduced from arXiv: 2604.13384 by Dongkuan Xu, Hengxu Li, Mingzhe Chen, Yuchen Liu.

Figure 1
Figure 1. Figure 1: The agentic control loop consists of rApp [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The implementation of agentic control stack with LLM and IaC. The LLM side performs intent [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Emergency phase percentile performance. TABLE II: Tail robustness (0–300 s) and emergency fairness. Metric Baseline A1GENT ∆ All-UE DL p05 (Mbps) 0.000 0.101 +0.101 Frac. samples < 0.10 Mbps (outage) 6.4% 4.9% −1.5 pp Frac. samples < 0.50 Mbps 44.4% 39.6% −4.8 pp Emergency p90/p10 ratio 7.96 3.21 −59.7% reallocate radio resources. Specifically, per-UE PDCP down￾link throughput at p10 increases from 0.280 t… view at source ↗
Figure 4
Figure 4. Figure 4: Affected UE timeline across phases. TABLE III: Recovery DL scheduler throughput (Mbps). Focus Cell Baseline mean Ours mean Median: Baseline→Ours 1 1.62 2.52 0.00→2.47 2 3.78 4.19 3.09→3.40 3 3.64 3.66 3.09→3.09 9 2.29 4.45 0.00→2.47 2) Recovery: sustained QoE and spatial rebalance: During the recovery phase, user experience continues to improve as network balance is restored. The 20th-percentile downlink t… view at source ↗
read the original abstract

Large language models (LLMs) open new possibilities for agentic control in Open RAN, allowing operators to express intents in natural language while delegating low-level execution to autonomous agents. We present A1gent, an agentic RAN control stack that decouples reasoning from real-time actuation. A non-RT agentic rApp compiles operator goals into typed A1 policy instances, and three task-oriented near-RT agentic xApps enforce them through a deterministic loop with plane-scoped actuation - E2 for mobility and load steering, and O1 for energy orchestration. This agentic reasoning-execution split ensures auditable coordination between RAN intelligent controller (RIC) tiers, supported by encoded guardrails and a fixed-priority action merger for conflict governance. A training-free adaptive policy tuner then refines bounded parameters using KPI memory without retraining, sustaining predictable adaptation. By integrating intent-driven planning with deterministic near-RT execution, A1gent advances Open RAN toward verifiable, self-governing, and reproducible agentic intelligence.

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 presents A1gent, an agentic RAN control stack that decouples non-RT LLM-based intent compilation into typed A1 policies from near-RT deterministic enforcement by three task-oriented xApps (via E2 for mobility/load steering and O1 for energy), supported by encoded guardrails, fixed-priority action merger for conflict resolution, and a training-free adaptive policy tuner that refines parameters from KPI memory.

Significance. If the described split and mechanisms can be shown to deliver the claimed properties, the work could meaningfully advance Open RAN by enabling natural-language operator intents while mitigating LLM nondeterminism through deterministic near-RT actuation and auditability features. The training-free tuner and plane-scoped actuation are potentially useful design elements for reproducible adaptation in production networks.

major comments (1)
  1. [Abstract] Abstract and high-level architecture description: the central claim that the reasoning-execution split 'ensures auditable coordination' and 'advances Open RAN toward verifiable, self-governing, and reproducible agentic intelligence' rests entirely on unshown mechanisms; the manuscript supplies no quantitative evaluation, simulation results, latency measurements, safety-violation analysis, or formal argument establishing that LLM-to-A1 translation succeeds at scale or that the actuation loop meets near-RT timing bounds under realistic traffic loads. This absence makes reliability and safety properties design goals rather than demonstrated outcomes and is load-bearing for the paper's thesis.
minor comments (2)
  1. The acronym 'A1gent' and the precise mapping from natural-language intents to typed A1 policy instances would benefit from an explicit definition and example on first use to improve readability for readers unfamiliar with the A1 interface.
  2. Consider adding a diagram or table that enumerates the three near-RT xApps, their E2/O1 scopes, and the priority rules in the action merger to make the conflict-governance mechanism concrete.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the thorough and constructive review. The major comment correctly identifies a gap in empirical validation for the central claims. We address this point below and will revise the manuscript to strengthen the evidence.

read point-by-point responses

  1. Referee: [Abstract] Abstract and high-level architecture description: the central claim that the reasoning-execution split 'ensures auditable coordination' and 'advances Open RAN toward verifiable, self-governing, and reproducible agentic intelligence' rests entirely on unshown mechanisms; the manuscript supplies no quantitative evaluation, simulation results, latency measurements, safety-violation analysis, or formal argument establishing that LLM-to-A1 translation succeeds at scale or that the actuation loop meets near-RT timing bounds under realistic traffic loads. This absence makes reliability and safety properties design goals rather than demonstrated outcomes and is load-bearing for the paper's thesis.

    Authors: We agree that the current manuscript is primarily an architectural framework paper and does not yet contain quantitative evaluations, latency measurements, safety-violation analysis, or formal arguments to fully substantiate the reliability and safety claims. The reasoning-execution split is presented as a design that enables auditable coordination through typed A1 policies, encoded guardrails, and deterministic near-RT xApps (detailed in Sections 3-5), but these mechanisms are described rather than empirically validated at scale. In the revised version we will add a dedicated Evaluation section that includes: (i) ns-3-based simulations of the E2/O1 actuation loop under realistic traffic loads to report end-to-end latency and compliance with near-RT timing bounds; (ii) quantitative assessment of guardrail effectiveness and conflict resolution success rate when conflicting intents are injected; (iii) accuracy metrics for LLM-to-A1 policy translation across a corpus of operator intents; and (iv) a concise formal argument (in the form of an invariant sketch) showing why the typed-policy plus guardrail design yields auditability. These additions will convert the stated properties from design goals into demonstrated outcomes supported by data. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper presents A1gent as an architectural framework decoupling non-RT LLM-based intent compilation from near-RT deterministic xApp enforcement via guardrails, priority-based conflict resolution, and a training-free adaptive policy tuner. No equations, fitted parameters, or derivation chains appear in the provided text. The central claim that this split advances Open RAN toward verifiable agentic intelligence is asserted directly from the described design components rather than reduced to self-citations, redefinitions, or renamed empirical patterns. The work is self-contained as a high-level system proposal with no load-bearing steps that equate outputs to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no equations, parameters, or explicit assumptions; ledger entries cannot be populated.

pith-pipeline@v0.9.0 · 5484 in / 1153 out tokens · 28196 ms · 2026-05-10T12:51:04.029287+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

12 extracted references · 12 canonical work pages

  1. [1]

    On AI Verification in Open RAN,

    R. Soundrarajan, C. Fiandrino, M. Polese, S. D’Oro, L. Bonati, and T. Melodia, “On AI Verification in Open RAN,”IEEE Communications Magazine, early access, 2025, doi: 10.1109/MCOM.001.2500318

    work page doi:10.1109/mcom.001.2500318 2025

  2. [2]

    Adaptive traffic steer- ing in Open RAN: Integrating rule-based policies with reinforcement learning,

    U. Sharma, H. Wei, M. Chen, J. Xu, and Y . Liu, “Adaptive traffic steer- ing in Open RAN: Integrating rule-based policies with reinforcement learning,” inProc. IEEE INFOCOM Workshops, 2025

    work page 2025

  3. [3]

    A comprehensive tutorial and survey of O-RAN: Exploring slicing-aware architecture, deployment options, use cases, and challenges,

    K. Alam et al., “A comprehensive tutorial and survey of O-RAN: Exploring slicing-aware architecture, deployment options, use cases, and challenges,”IEEE Commun. Surv. Tutor., vol. 28, pp. 1637–1678, 2026, doi: 10.1109/COMST.2025.3598406

    work page doi:10.1109/comst.2025.3598406 2026

  4. [4]

    O-RAN xApps: Survey and research challenges,

    A. Elyasi, A. Ashdown, K. M. Rumman, and F. Restuccia, “O-RAN xApps: Survey and research challenges,”Comput. Netw., 2025

    work page 2025

  5. [5]

    Open RAN testbeds with controlled air mobil- ity,

    M. Mushi et al., “Open RAN testbeds with controlled air mobil- ity,”Computer Communications, vol. 228, p. 107955, 2024, doi: 10.1016/j.comcom.2024.107955

    work page doi:10.1016/j.comcom.2024.107955 2024

  6. [6]

    Agentran: An agentic ai architecture for autonomous control of open 6g networks,

    M. Elkael et al., “AgentRAN: An agentic AI architecture for autonomous control of open 6G networks,” arXiv:2508.17778 [cs.AI], 2025

    work page arXiv 2025

  7. [7]

    Demo: human-in-the-loop agentic reconfiguration of edge 5G networks via dual-MCP and LLM reasoning,

    E. Baena, A. Mandal, and D. Koutsonikolas, “Demo: human-in-the-loop agentic reconfiguration of edge 5G networks via dual-MCP and LLM reasoning,” inProc. ACM MobiHoc, Houston, TX, USA, 2025, pp. 391– 392, doi: 10.1145/3704413.3765314

    work page doi:10.1145/3704413.3765314 2025

  8. [8]

    Edge agentic AI framework for autonomous network optimisation in O-RAN,

    A. Salama, Z. Nezami, M. M. H. Qazzaz, M. Hafeez, and S. A. R. Zaidi, “Edge agentic AI framework for autonomous network optimisation in O-RAN,” inProc. IEEE Int. Symp. Pers., Indoor and Mobile Radio Commun. (PIMRC) Workshops, 2025

    work page 2025

  9. [9]

    ALL- STaR: Automated LLM-driven scheduler generation and testing for intent-based RAN,

    M. Elkael, M. Polese, R. Prasad, S. Maxenti, and T. Melodia, “ALL- STaR: Automated LLM-driven scheduler generation and testing for intent-based RAN,” arXiv:2505.18389 [cs.NI], 2025

    work page arXiv 2025

  10. [10]

    AutoRAN: Automated and Zero-Touch Open RAN Systems,

    S. Maxenti et al., “AutoRAN: Automated and zero-touch open RAN systems,” arXiv:2504.11233 [cs.NI], 2025

    work page arXiv 2025

  11. [11]

    PACIFISTA: Conflict evaluation and man- agement in Open RAN,

    P. Brach del Prever et al., “PACIFISTA: Conflict evaluation and man- agement in Open RAN,”IEEE Trans. Mobile Comput., vol. 24, no. 10, pp. 10590–10605, 2025, doi: 10.1109/TMC.2025.3570632

    work page doi:10.1109/tmc.2025.3570632 2025

  12. [12]

    O-RAN with machine learning in ns-3,

    W. Garey, R. A. Rouil, E. Black, T. Ropitault, and W. Gao, “O-RAN with machine learning in ns-3,” inProc. Workshop on ns-3 (WNS3), Ballston, V A, USA, 2023, pp. 60–68, doi: 10.1145/3592149.3592157

    work page doi:10.1145/3592149.3592157 2023