Agentic AI-Based Joint Computing and Networking via Mixture of Experts and Large Language Models

Robert-Jeron Reifert; Alaa Alameer Ahmad; Hayssam Dahrouj; Aydin Sezgin

cs.LGApr 7, 2026

Agentic AI-Based Joint Computing and Networking via Mixture of Experts and Large Language Models

Authors: Robert-Jeron Reifert, Alaa Alameer Ahmad, Hayssam Dahrouj, Aydin Sezgin

Organizations: Digital Communication Systems, Ruhr University Bochum, Bochum, Germany · Cariad SE, Wolfsburg, Germany · Department of Electrical Engineering, University of Sharjah, Sharjah, United Arab Emirates

Abstract

Future sixth-generation (6G) mobile networks are envisioned to be equipped with a diverse set of powerful, yet highly specialized, optimization experts. Such a promising vision is concurrently expected to give rise to the need for scalable mechanisms that can select, combine, and orchestrate such experts based on high-level intent and uncertainty descriptions. In this paper, we propose an agentic artificial intelligence (AI)-based network optimization framework that integrates mixture of experts (MoE) architectures with large language models (LLMs). Under the proposed framework, the employed LLM acts as a semantic gate to reason over operator objectives and dynamically compose suitable optimization agents. The proposed framework is formulated in a model-agnostic manner and bridges human-readable network intents with low-level resource allocation decisions, enabling flexible optimization across heterogeneous objectives and operating conditions. As a representative instantiation, we apply the framework to a joint communication and computing network and design a library of specialized optimization experts covering throughput, fairness, and delay-driven objectives under both regular and robust conditions. Numerical simulations demonstrate that the proposed agentic MoE framework consistently achieves near-optimal performance compared to exhaustive expert combinations while outperforming individual experts across diverse objectives, including delay minimization and throughput maximization.

Figures & tables

Fig. 1: General agentic AI-based network optimization framework with an LLM-enabled gate and a library of deep learning-based optimization experts, illustrating the interaction between network operator intent and network infrastructure.

System prompt
	You are functioning as a gate network that receives a query from the network operator and your task is to route the question to a subset of optimization expert from a library. You have $A$ experts that can be used to resolve queries either by their own or on combinations. This depends on the query if the requested information need to be a combination of the results from several experts or one expert can fully address the query. This can be solely determined by the description of each expert area of expertise. Here is a detailed description of the available experts and their area of specialization:
	-Agent description (including utility function, goal, operating conditions, parameter set, etc.)-
Operator query
	-Network setup description (e.g., XR, computing, imperfect channels, etc.)-
	-Operator goal description (e.g., sum-rate maximization, etc.)-

TABLE I: High-level illustration of system prompt

\mathcal{Q}^{\text{sys}}

and operator query

\mathcal{Q}^{\text{op}}(\mathcal{S})

Fig. 2: Joint communication and computing network consisting of a multi-antenna BS with embedded computing capabilities serving multiple single-antenna devices.

Idx.	Name	Role	System Impact	$\mathbf{x}_{a}$	$f_{\text{utility},a}(\mathbf{x}_{a})$	$\mathcal{C}_{a}(\mathbf{x}_{a})$	$\lambda_{\text{utility},a}$
$1$	Comm_SumR_Reg	Comm. sum rate max.	Throughput	$\mathbf{p}^{\text{tx}}$	$\sum_{k\in\mathcal{K}}r_{k}^{\text{tx}}$	$\mathcal{D}_{\text{comm}}\cup\mathcal{P}_{\text{comm}}$	–
$2$	Comm_SumR_Rob	Comm. sum rate max.	Throughput	$\mathbf{p}^{\text{tx}}$	$\lambda_{\text{sumR},a}^{\gamma}$	$\mathcal{D}_{\text{comm}}\cup\mathcal{P}_{\text{comm}}\cup\mathcal{R}_{\text{comm}}$	$\sum_{k\in\mathcal{K}}r_{k}^{\text{tx}}$
$3$	Comp_SumR_Reg	Comp. sum rate max.	Throughput	$p^{\text{co}},\mathbf{f}^{\text{co}}$	$\sum_{k\in\mathcal{K}}r_{k}^{\text{co}}$	$\mathcal{D}_{\text{comp}}\cup\mathcal{P}_{\text{comp}}$	–
$4$	Comp_SumR_Rob	Comp. sum rate max.	Throughput	$p^{\text{co}},\mathbf{f}^{\text{co}}$	$\lambda_{\text{sumR},a}^{\gamma}$	$\mathcal{D}_{\text{comp}}\cup\mathcal{P}_{\text{comp}}\cup\mathcal{R}_{\text{comp}}$	$\sum_{k\in\mathcal{K}}r_{k}^{\text{co}}$
$5$	JCC_SumR_Reg	Sum rate max.	Throughput	$\mathbf{p}^{\text{tx}},p^{\text{co}},\mathbf{f}^{\text{co}}$	$\sum_{k\in\mathcal{K}}(r_{k}^{\text{tx}}+r_{k}^{\text{co}})$	$\mathcal{D}_{\text{comp}}\cup\mathcal{P}_{\text{comp}}$	–
$6$	JCC_SumR_Rob	Sum rate max.	Throughput	$\mathbf{p}^{\text{tx}},p^{\text{co}},\mathbf{f}^{\text{co}}$	$\lambda_{\text{sumR},a}^{\gamma}$	$\mathcal{D}_{\text{joint}}\cup\mathcal{P}_{\text{joint}}\cup\mathcal{R}_{\text{joint}}$	$\sum_{k\in\mathcal{K}}(r_{k}^{\text{tx}}+r_{k}^{\text{co}})$

TABLE II: List of available experts with index

a\in\mathcal{N}

, name, role, and system impact, as well as the corresponding optimization formulation in terms of variables (

\mathbf{x}_{a}

), objective function (

f_{\text{utility},a}(\mathbf{x}_{a})

), constraint set (

\mathcal{C}_{a}(\mathbf{x}_{a})

), and robustness utility (

\lambda_{\text{utility},a}

Fig. 3: Generic DNN structure and uncertainty injection mechanism utilized in the proposed case study.

Network resources	DNN parameters
$L=4$ antennas, $K=4$ users	$\eta_{n}^{h}=10$ layers (ReLU)
$P^{\text{max}}=34$ dBm	$\chi_{n}^{h}=400$ neurons per layer
$W=5$ MHz, $\Gamma^{\text{gap}}=9.5$ dB	$500$ epochs, $50$ minibatches
$\sigma^{2}=-75$ dBm/Hz	$1000$ batch & $2000$ validation size
$F^{\text{max}}=4.6$ GHz, $\tau=10^{-28}$ , $\mu=3$	$4000$ testing size
$D_{k}^{\text{out}}=2.5\cdot 10^{4}$ bit, $D_{k}^{\text{in}}=5\cdot 10^{4}$ bit

TABLE III: Simulation parameters.

Framework setup
Query	I have connected extended reality devices. I am rendering their field of view and then transmitting the video stream. All users should be treated fair and have minimal latency
Expert library	Experts $1$ - $4$ , $7$ - $10$ , $13$ - $16$ , and $19$ - $22$
Optimal benchmarks	Experts $5$ , $6$ , $11$ , $12$ , $17$ , $18$ , $23$ , and $24$
Agentic AI solution
Interpretation	To address your query, we need to focus on minimizing the maximum delay for both communication and computing, ensuring fairness among all users. Since you are dealing with both rendering (computing) and video transmission (communication), we should consider a joint approach. Given that you want minimal latency and fairness, the experts for minimizing maximum delay under regular conditions for both communication and computing should be used.
Selection	$19$ Comm_MaxT_Reg & $21$ Comp_MaxT_Reg

TABLE IV: Description of simulation Set 1.

Fig. 4: Experts, benchmarks, and agentic AI solution space as scatter plot, including infeasible results for simulation Set 1.

Fig. 5: Experts, benchmarks, and agentic AI solution space as bar plot, including infeasible results for simulation Set 1.

Fig. 6: Expert combinations and agentic AI solution space as scatter plot, including infeasible results for simulation Set 1.

Fig. 7: Expert combinations and agentic AI solution space as bar plot, including infeasible results for simulation Set 1.

Set		End-to-end	Netw. delay	Inference	Accuracy
1	avg	$9.363$ s	$1.09$ s	$0.177$ s	$90\%$
2	avg	$5.926$ s	$1.192$ s	$0.176$ s	$100\%$
3	avg	$7.817$ s	$1.5$ s	$0.182$ s	$100\%$

TABLE V: Runtime comparison across all simulation sets (

10

independent trials, each), with end-to-end time, average network delay, expert inference time, and overall accuracy.

Framework setup
Query	I want to jointly achieve optimal network throughput and fairness for a robust extended reality scenario
Expert library	Experts $1$ - $12$
Optimal benchmarks	Experts $13$ - $18$
Agentic AI solution
Interpretation	To jointly achieve optimal network throughput and fairness for a robust extended reality scenario, we need to consider both communication and computing aspects under impaired conditions. This requires a combination of experts that focus on robust maximization of sum-rates and robust maximization of minimum rates.
Selection	$6$ JCC_SumR_Rob & $12$ JCC_MinR_Rob

TABLE VI: Description of simulation Set 2.

Fig. 8: Experts, benchmarks, and agentic AI solution space as scatter plot, including infeasible results for simulation Set 2.

Fig. 9: Experts, benchmarks, and agentic AI solution space as bar plot showing the sum-rate, including infeasible results for simulation Set 2.

Fig. 10: Experts, benchmarks, and agentic AI solution space as bar plot showing the min-rate, including infeasible results for simulation Set 2.

Fig. 11: Expert combinations and agentic AI solution space as scatter plot, including infeasible results for simulation Set 2.

Framework setup
Query	Computing requirements and channels are subject to estimation errors. I want to serve all users with an equally good delay but also achieve optimal network throughput in terms of data processing and communication
Expert library	Experts $1$ - $30$
Optimal benchmarks	—-
Agentic AI solution
Interpretation	To address your query, we need to focus on both fairness in terms of delay and optimal throughput under conditions where both computing requirements and channels are subject to estimation errors. This involves a combination of experts that handle robust conditions for both delay fairness and throughput optimization.
Selection	$24$ JCC_MaxT_Rob & $6$ JCC_SumR_Rob

TABLE VII: Description of simulation Set 3.

Fig. 12: Library experts and agentic AI solution space as scatter plot for simulation Set 3.

Fig. 13: Experts and agentic AI solution space as bar plot showing the robust sum-rate for simulation Set 3.

Fig. 14: Experts and agentic AI solution space as bar plot showing the robust log-max-delay, including infeasible results for simulation Set 3.

Fig. 15: Expert combinations and agentic AI solution space as scatter plot for simulation Set 3.

Successful Examples
Query	I have very uncertain network estimates and want to avoid extreme delays for any user. However, throughput should not collapse completely. The network must remain resilient overall.
Solution	JCC_MaxT_Rob , $\alpha_{24}=0.6$ & JCC_SumR_Rob , $\alpha_{6}=0.4$
Query	I want to ensure that critical communication links achieve guaranteed minimum rates while optimizing the total throughput for the rest under regular conditions.
Solution	Comm_MinR_Reg , $\alpha_{7}=0.5$ & Comm_SumR_Reg , $\alpha_{1}=0.5$
Query	Channel estimates are almost perfect, but computing workloads fluctuate unpredictably. The main objective is to guarantee fairness in computing performance while maintaining reasonable throughput.
Solution	Comp_MinR_Rob , $\alpha_{10}=0.6$ & Comm_SumR_Reg , $\alpha_{1}=0.4$

TABLE VIII: Example queries, LLM-selected experts, and accuracy in terms of human-expert solution.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Fig. 16: Visual characterization of experts from Table II .

General setup
	You are a helpful assistant that receive requests in natural language and then, based on the provided context you must choose one or a combination of suitable expert tools which can resolve the query asked. You operate as an intelligent assistant embedded within a wireless network operator environment. You are functioning as a router/gate network that receives a query or a question from the network operator and your task is to route the question to an optimization expert.
Expert setup
	You have 30 experts that can be used to resolve queries either by their own or on combinations. This depends on the query if the requested information need to be a combination of the results from several experts or one expert can fully address the query. This can be solely determined by the description of each expert area of expertise. Here is a detailed description of the available experts and their area of specialization:
	1) Comm_SumR_Reg: Expert for maximization of sum-communication-rates of all users in network. Communication throughput focused. Optimized solution accounts for regular conditions, with accurate channel estimations and perfectly known channel state information.
	[ … ]
Tool setup

Appendix

TABLE IX: Excerpt of the system prompt

\mathcal{Q}^{\text{sys}}

used for the LLM-enabled gate network.

LLM Agent Orchestration

Large language model agent orchestration coordinates models, tools, and workflows to complete complex tasks reliably and efficiently.

See all 458 papers

6G Needs Agents: Toward Agentic AI-Native Networks for Autonomous Intelligence
May 2, 2026Mohamed Amine Ferrag, Abderrahmane Lakas, Merouane Debbah6G Networks LLM Agent Orchestration
Department of Computer and Network Engineering, United Arab Emirates University, UAE · Research Institute for Digital Future, Khalifa University, UAE
LLM-Powered Agentic AI for 5G/6G Networks: A Tutorial and Survey on Architectures, Protocols, and Standardization
Jul 17, 2026Mazene Ameur, Abdelkader Mekrache, Bouziane Brik +1Multi-Agent LLM Systems Agent Communication Protocols
EURECOM, France · Simula Metropolitan, Norway · University of Sharjah, UAE
Agentic Wireless Communication for 6G: Intent-Aware and Continuously Evolving Physical-Layer Intelligence
Feb 19, 2026Zhaoyang Li, Xingzhi Jin, Junyu Pan +2Wireless Communications 6G Networks
College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou, China.

May 2, 2026J/K move · Enter open · S save

6G Needs Agents: Toward Agentic AI-Native Networks for Autonomous Intelligence

Mohamed Amine Ferrag, Abderrahmane Lakas, Merouane Debbah

Department of Computer and Network Engineering, United Arab Emirates University, UAE · Research Institute for Digital Future, Khalifa University, UAE

6G Networks LLM Agent Orchestration Wireless Communications Multi-Agent LLMs

Open paper AskPDF arXiv

Sixth-generation (6G) networks are increasingly envisioned as AI-native infrastructures integrating communication, sensing, and computing into a unified fabric. However, existing approaches remain largely optimization-centric, relying on closed-loop control with limited reasoning capability. In this paper, we argue for a paradigm shift toward Agentic AI-Native 6G, in which Large Language Model (LLM)-based agents operate as bounded, policy-governed reasoning entities within a semantic control plane layered above deterministic 3GPP infrastructure. We propose a four-layer architecture that integrates deterministic network infrastructure, semantic abstraction of intent and context, hierarchical reasoning, and a distributed multi-agent fabric spanning device, edge, and core domains. To assess feasibility, we develop a proof-of-concept agentic reasoning and orchestration framework and conduct an extensive empirical study using a domain-specific 6G benchmark under realistic deployment constraints. Our results reveal a fundamental tradeoff between reasoning capability and system efficiency, showing that no single model simultaneously satisfies latency, throughput, and accuracy requirements. Instead, heterogeneous deployment of LLM agents across the device--edge--core continuum is necessary to balance these constraints. We further demonstrate that quantization introduces non-uniform effects across models, reinforcing the need for system-level optimization rather than model-level compression alone. These findings establish agentic intelligence as a viable architectural direction for 6G and highlight key challenges in achieving scalable, trustworthy, and self-reasoning networks. All experimental results and evaluation scripts are publicly available to support reproducibility.

Abstract

LLM Agent Orchestration

Explore similar work

6G Needs Agents: Toward Agentic AI-Native Networks for Autonomous Intelligence

LLM-Powered Agentic AI for 5G/6G Networks: A Tutorial and Survey on Architectures, Protocols, and Standardization

Agentic Wireless Communication for 6G: Intent-Aware and Continuously Evolving Physical-Layer Intelligence