PaperlayerView exact revision on arXiv

Source-linked AI summary

ComAgent: Multi-LLM based Agentic AI Empowered Intelligent Wireless Networks

Haoyun Li, Ming Xiao, Kezhi Wang, Robert Schober, Dong In Kim, Yong Liang Guan

arXiv:2601.19607v1cs.AI

TL;DR

Complex, coupled 6G wireless optimization still requires extensive manual modeling and algorithm design. ComAgent coordinates specialized LLM agents in a closed-loop workflow, achieving 100% problem formulation and a 72% solution-solved rate across tasks.

  • Problem

    Emerging wireless networks create high-dimensional, non-convex, multi-objective optimization problems that require substantial expert effort to formulate and solve.

  • Method

    ComAgent coordinates specialized agents in a recursive Perception-Planning-Action-Reflection cycle to generate mathematical formulations, executable code, and physics-aware validation.

  • Results

    The framework achieved a 100% Problem Formulation Rate and a 72% Solution Solved Rate, outperforming the single-LLM baseline across evaluated tasks.

  • Takeaways & Limitations

    ComAgent demonstrates a multi-LLM architecture that automates wireless optimization workflows while reaching solution quality comparable to expert-designed baselines.

  • Takeaways & Limitations

    The framework currently automates specific wireless optimization tasks rather than controlling dynamic real-world wireless networks.

Abstract

from arXiv · show

Emerging 6G networks rely on complex cross-layer optimization, yet manually translating high-level intents into mathematical formulations remains a bottleneck. While Large Language Models (LLMs) offer promise, monolithic approaches often lack sufficient domain grounding, constraint awareness, and verification capabilities. To address this, we present ComAgent, a multi-LLM agentic AI framework. ComAgent employs a closed-loop Perception-Planning-Action-Reflection cycle, coordinating specialized agents for literature search, coding, and scoring to autonomously generate solver-ready formulations and reproducible simulations. By iteratively decomposing problems and self-correcting errors, the framework effectively bridges the gap between user intent and execution. Evaluations demonstrate that ComAgent achieves expert-comparable performance in complex beamforming optimization and outperforms monolithic LLMs across diverse wireless tasks, highlighting its potential for automating design in emerging wireless networks.

I. INTRODUCTION

Emerging 6G networks create heterogeneous, intent-driven optimization demands that challenge conventional manually modeled workflows and monolithic LLM approaches. ComAgent addresses this gap through a multi-LLM, tool-augmented closed-loop system that connects intent understanding, mathematical formulation, and code-based validation.

  • I. INTRODUCTION: Intent-driven 6G applications impose stringent and often conflicting requirements on heterogeneous infrastructures and network functions.The context includes SAGIN, ultra-massive MIMO, RIS, SWIPT, ISAC, edge intelligence, XR, autonomous systems, and large-scale IoT.
  • I. INTRODUCTION: Traditional optimization workflows require manually designed models and problem-specific algorithms for high-dimensional, strongly non-convex, and multi-objective problems.These workflows exploit problem structures for decomposition and non-convexity transformation, but real-world complexity makes such optimization challenging.
  • I. INTRODUCTION: LLMs can interpret high-level textual specifications and unify configuration generation, optimization, and troubleshooting through zero- or few-shot adaptation.This enables natural-language directives to drive network-configuration and control-algorithm synthesis, but monolithic integration into wireless optimization remains limited.
  • I. INTRODUCTION: Agentic AI extends LLM assistance with iterative task decomposition, explicit planning, tool invocation, and reflection based on execution feedback.The tool-augmented, feedback-driven loop is presented as particularly aligned with wireless optimization.
  • I. INTRODUCTION: ComAgent coordinates Literature, Planning, Coding, and Scoring Agents in a recursive cognitive cycle for autonomous end-to-end wireless optimization.The framework is positioned as a closed-loop system bridging intent understanding, mathematical rigor, and code-centric validation.

II. THE LANDSCAPE OF LLM-DRIVEN OPTIMIZATION: PROGRESS AND GAPS · A. Recent advances in LLM-enabled wireless optimization · B. Open challenges toward end-to-end agentic optimization copilots

LLMs enable semantic-aware wireless network management but remain inadequate as standalone numerical optimizers for constrained, non-convex problems. Existing hybrid and multi-agent approaches leave gaps in autonomous workflow redesign, rigorous formulation, solver execution, and verification, motivating ComAgent’s end-to-end multi-LLM architecture.

  • A. Recent advances in LLM-enabled wireless optimization: LLMs shift intelligent network management from rigid data-driven deep-learning pipelines toward flexible, semantic-aware control.
  • A. Recent advances in LLM-enabled wireless optimization: Monolithic LLMs struggle with high-precision, non-convex wireless optimization because they lack native numerical optimization, convergence guarantees, and complex-constraint satisfaction mechanisms.
  • A. Recent advances in LLM-enabled wireless optimization: Hybrid frameworks use LLMs as heuristic or black-box components inside human-designed pipelines, while conventional methods handle numerical beamforming optimization.Reported methods include ADMM, MM, and FP, alongside LLM-based user-association solving.
  • A. Recent advances in LLM-enabled wireless optimization: Multi-agent systems decompose network-management tasks by domain, assigning specialized agents to functions such as traffic, QoS, security, threat assessment, and decision support.
  • B. Open challenges toward end-to-end agentic optimization copilots: Hybrid schemes constrain LLMs to rigid, predetermined pipelines, preventing autonomous workflow reorganization, solution-strategy redesign, and feasibility validation when contexts change.
  • B. Open challenges toward end-to-end agentic optimization copilots: ComAgent proposes a multi-LLM agentic architecture that autonomously manages the full optimization workflow through specialized agents and feedback-driven coordination.The framework is presented as bridging manual design exploration with autonomous, code-centric validation for dynamic wireless networks.
  • B. Open challenges toward end-to-end agentic optimization copilots: Existing multi-agent systems emphasize high-level automation rather than translating intent into mathematically rigorous, executable solvers under physical constraints.
  • B. Open challenges toward end-to-end agentic optimization copilots: Generative formulation agents paired with separate deep-RL solvers reduce the gap but retain structural disconnects between formulation and solution generation and lack unified execution-grounded verification.

III. AGENTIC AI FRAMEWORK AND METHODOLOGY · A. A communication-oriented agentic AI architecture

ComAgent is a communication-oriented multi-LLM optimization copilot that replaces stand-alone generation with a recursive Perception–Planning–Action–Reflection architecture. The framework grounds wireless-network design in explicit intents, engineering states, executable validation, and physics-aware iterative refinement.

  • III. AGENTIC AI FRAMEWORK AND METHODOLOGY: ComAgent is tailored for wireless optimization, producing numerically precise, constraint-intensive, and physically meaningful decisions across emerging network settings.The architecture targets scenarios including ISAC, SWIPT, RIS, SAGIN, low-altitude networks, and edge intelligence.
  • A. A communication-oriented agentic AI architecture: The communication-oriented controller recursively cycles through Perception, Planning, Action, and Reflection from problem interpretation to solver execution and feasibility checking.This lifecycle distinguishes the framework from conventional static LLM pipelines.
  • A. A communication-oriented agentic AI architecture: Perception continuously combines semantic intents, such as QoS and latency targets, with engineering states including CSI, topology, resource budgets, and protocol constraints.It operates as an ongoing state-awareness interface rather than a one-time initialization step.
  • A. A communication-oriented agentic AI architecture: At macro and micro levels, perception parses intent into design constraints while specialized agents interpret retrieval, solver, numerical, and simulation feedback.Examples include beamforming objectives, power allocation requirements, energy guarantees, solver infeasibility, and simulation mismatches.
  • A. A communication-oriented agentic AI architecture: Planning recursively decomposes wireless-design intent into a solvable and verifiable pipeline spanning modeling, variable and constraint specification, algorithm selection, and baseline construction.Planning therefore produces more than a formal objective by orchestrating multiple stages of optimization workflow.
  • A. A communication-oriented agentic AI architecture: Action converts the LLM into an active executor through literature retrieval, solver calls, dataset generation, simulation execution, and baseline reproduction.Executable validation under physical constraints establishes the correctness of wireless optimization.
  • A. A communication-oriented agentic AI architecture: Reflection operates as a physics-aware critic that detects hallucinations, logical inconsistencies, and feasibility violations before iteratively refining strategies and actions.It checks constraint satisfaction, physical meaningfulness, rank conditions, and power and energy requirements, while verifying code fidelity.

B. End-to-end workflow instantiation for network optimization

The framework is instantiated as a four-stage workflow aligned with wireless network design practices. It targets solver-ready models and execution-verified results under communication constraints, rather than merely plausible descriptions.

  • The workflow comprises knowledge acquisition, plan formulation, data preparation, and simulation generation.
  • Its objective is to produce solver-ready models and execution-verified results under communication constraints, not only plausible descriptions.

1) Knowledge acquisition: · 2) Plan formulation:

ComAgent grounds wireless optimization requests by retrieving and refining relevant literature, then converts the resulting knowledge into an implementation-oriented, solver-ready plan. The plan iteratively specifies models, objectives, constraints, solution routes, and linked execution subtasks through instructor-guided review.

  • 1) Knowledge acquisition:: The Literature Agent parses each natural-language request to identify the task type, evaluation targets, and essential modeling elements governing feasibility and performance.These elements include channel assumptions, antenna settings, hardware constraints, and task-specific metrics for tasks such as SWIPT beamforming, ISAC co-design, RIS configuration, and SAGIN resource management.
  • 1) Knowledge acquisition:: It uses Semantic Scholar and web browsers to iteratively query, inspect, and filter candidate papers and baselines according to scenario alignment and constraints.The process avoids a single search-and-summarize step and accounts for modeling differences that can affect wireless optimization.
  • 1) Knowledge acquisition:: The Literature Agent ranks selected papers by scenario relevance and extracts formulations, constraints, and solution methods into System Memory for the next stage.The stored knowledge is combined with the original user query.
  • 2) Plan formulation:: Stage II translates the grounded query and literature knowledge into a solver-ready optimization plan under a Planning Instructor’s coordination.The agents use chain-of-thought, reasoning-action, and plan-and-solve prompting strategies to enhance reasoning for complex tasks.
  • 2) Plan formulation:: The Planning Agent specifies implementable system models, design objectives, optimization variables, constraints, performance metrics, and a solution route suited to the problem’s structure and numerical properties.Practical requirements include power budgets, SINR guarantees, sensing accuracy targets, and energy harvesting thresholds.
  • 2) Plan formulation:: The Planning Instructor reviews drafts for missing constraints, ambiguous definitions, and difficult implementation steps, prompting revisions until the plan is internally consistent.This review-and-revision loop makes the planning process implementation-oriented.
  • 2) Plan formulation:: The final output is a structured execution chain linking formulation, algorithm design, baseline construction, and experiment settings for downstream execution.Fig. 2 illustrates this autonomous workflow through a MIMO SWIPT beamforming optimization case study spanning four stages.

3) Data preparation: · 4) Simulation generation:

Data preparation aligns generated wireless data with the finalized plan through supervised coding and iterative consistency checks. Simulation generation then produces modular executable code and validates it through execution, wireless-constraint checks, scoring, and iterative refinement.

  • 3) Data preparation:: The Data Instructor retrieves the user query and finalized plan, then supervises the Coding Agent in generating or collecting model-consistent wireless data.The data includes channel realizations, topology generation, parameter settings, and task-specific data when needed.
  • 3) Data preparation:: Iterative checks verify code execution, dimensional consistency, path-loss trends, and power normalization before storing the validated dataset.When issues occur, the Coding Agent revises the script and repeats the process until the dataset matches the plan.
  • 4) Simulation generation:: Stage IV converts the solution plan and prepared dataset into executable sandbox simulations using modular system-model, solver, baseline, and metric components.The Coding Agent records implementation modifications to maintain alignment with the plan and improve traceability.
  • 4) Simulation generation:: Validation checks whether the implementation satisfies wireless constraints and whether its results are physically meaningful, beyond merely confirming code execution.The Scoring Agent conducts this communication-relevant validation through two feedback branches.
  • 4) Simulation generation:: Compilation or execution failures trigger specific error reporting, Coding Agent self-reflection, root-cause diagnosis, and corrective iteration.Examples include syntax errors and runtime exceptions.
  • 4) Simulation generation:: Successful simulations undergo feasibility, numerical-stability, constraint-satisfaction, and parameter-behavior checks, including rate versus transmit power.These checks assess whether key metrics behave as expected under parameter changes.
  • 4) Simulation generation:: 0 to 1 is the reward model’s quantitative score range, with higher scores indicating stronger alignment with task requirements and physical constraints.The Coding Agent uses the score and qualitative feedback to iteratively refine the implementation and maximize the reward.

IV. PERFORMANCE EVALUATION: FROM AUTONOMOUS DESIGN TO ROBUST GENERALIZATION · A. Case study 1: MIMO SWIPT beamforming optimization

The performance evaluation presents two case studies using Claude-4.5-Sonnet as the base model for all agents. The first case study tests autonomous orchestration on an expert-designed MIMO SWIPT beamforming optimization problem.

  • IV. PERFORMANCE EVALUATION: FROM AUTONOMOUS DESIGN TO ROBUST GENERALIZATION: Two case studies evaluate the proposed multi-LLM agentic AI system, with Claude-4.5-Sonnet serving as the base model for all agents.This establishes the evaluation scope and common model configuration.
  • A. Case study 1: MIMO SWIPT beamforming optimization: The first case study targets autonomous end-to-end orchestration for a non-trivial MIMO SWIPT beamforming optimization problem designed by a human expert.The task is framed as a test of the framework’s ability to translate and execute a complex wireless-network design problem.
  • A. Case study 1: MIMO SWIPT beamforming optimization: The scenario is a MIMO wireless downlink SWIPT network with K information receivers and J energy receivers under shadowed Rician fading channels.The network jointly serves information and energy receivers over shadowed Rician fading links.
  • A. Case study 1: MIMO SWIPT beamforming optimization: The optimization objective is to maximize the total sum-rate of all information receivers while guaranteeing each energy receiver’s minimum harvested energy.The task combines an information-throughput objective with harvested-energy constraints.
  • A. Case study 1: MIMO SWIPT beamforming optimization: The framework begins the MIMO SWIPT beamforming process from a user query expressed in natural language.The query initiates the autonomous design workflow for the beamforming task.
  • A. Case study 1: MIMO SWIPT beamforming optimization: The user query specifies one base station simultaneously serving multiple information and energy receivers, with ULAs at the base station, information receivers, and energy receivers.The query defines the antenna configuration and multi-receiver downlink setting.
  • A. Case study 1: MIMO SWIPT beamforming optimization: The query assumes shadowed Rician fading links and a linear radio-frequency energy-harvesting model for optimizing the base-station beamforming design.These assumptions provide the channel and harvesting-model conditions for the optimization.

1) Autonomous end-to-end solution: · 2) Performance verification:

ComAgent autonomously translates high-level MIMO SWIPT intents into mathematical formulations and simulation code through specialized agents and iterative verification. In performance verification, the generated framework models shadowed Rician fading and achieves average sum-rate performance comparable to traditional WMMSE beamforming.

  • 1) Autonomous end-to-end solution:: The Literature Agent retrieves and filters MIMO SWIPT references, while the Planning Agent translates user intent into a rigorous mathematical framework.The Planning Agent collaborates with a Planning Instructor rather than simply copying formulas.
  • 1) Autonomous end-to-end solution:: For the non-convex SWIPT beamforming problem, the Planning Agent sequences channel modeling, SDR and SCA transformations, and baseline selection.This decomposition forms a logical chain from problem modeling through algorithm design to comparative analysis.
  • 1) Autonomous end-to-end solution:: The Coding and Scoring Agents form a feedback loop in which generated simulations are evaluated for syntax correctness, physical feasibility, and semantic constraints.The Scoring Agent serves as an expert critic during execution.
  • 2) Performance verification:: The generated simulation code includes system parameters, algorithm design, and baseline implementations for K = 2 IRs and J = 2 ERs in a 30 m × 30 m region.The system uses a carrier frequency of fc = 28 GHz.
  • 2) Performance verification:: The verified setup uses M = 4 BS antennas at h = 5 m, with Nr = Ne = 3 antennas per IR and ER, half-wavelength spacing, and Rician fading with Krician = 6 dB.Large-scale attenuation uses L0 = 30 dB, α = 3.5, and σsh = 6 dB.
  • 2) Performance verification:: Average achievable sum-rate versus maximum transmit power shows agent-generated beamforming performance comparable to traditional WMMSE.The comparison indicates that the generated code implements iterative balancing of signal strength and inter-user interference.

B. Case study 2: Generalization and robustness across tasks

ComAgent was evaluated on 25 expert-designed wireless optimization tasks spanning classical resource allocation, NOMA, and RIS, against two single-LLM baselines. The baselines showed major weaknesses in formulation, execution, and robustness, while ComAgent achieved a 100% result on an unspecified metric in the provided passage.

  • Evaluation setup: The evaluation used 25 human-expert-designed optimization tasks spanning classical resource allocation, NOMA, and RIS, comparing ComAgent with two single-LLM baselines.The baselines were a single LLM with standard input-output prompting and a single LLM plus PS.
  • Baseline comparison: 0% Problem Formulation Rate, 24% Code Execution Rate, and 24% Solution Solved Rate characterized the single LLM baseline’s leap-to-code failure mode.Generated scripts were often syntactically incorrect or logically flawed, including cases that ignored physical power constraints.
  • Baseline comparison: 56% Formulation Rate and 56% Solution Solved Rate improved under Single LLM plus PS, but its 20% 1st-Try Success Rate and 2.44 Average Attempt Times indicated limited robustness.The static approach lacked feedback and struggled to recover from initial errors.

V. CHALLENGES AND FUTURE DIRECTIONS · A. From task-based execution to continuous operation · B. Balancing reasoning capabilities with real-time latency

ComAgent’s task-specific automation must evolve into an autonomous controller for dynamic wireless infrastructures. Future systems must combine persistent operation with hierarchical control that preserves advanced reasoning without violating real-time latency requirements.

  • V. CHALLENGES AND FUTURE DIRECTIONS: The framework currently automates specific wireless optimization tasks, but real-world deployment requires managing dynamic infrastructures beyond static problem solving.This transition introduces practical engineering challenges for next-generation agentic AI systems.
  • A. From task-based execution to continuous operation: The current one-shot workflow receives a query, executes a plan, and terminates after generating a solution.This episodic operation is effective for offline optimization but does not match persistent live-network demands.
  • A. From task-based execution to continuous operation: Live wireless networks require continuous service because objectives and decision contexts constantly evolve.Future architectures should maintain persistent state and continuously monitor network key performance indicators.
  • B. Balancing reasoning capabilities with real-time latency: Extensive inter-agent interactions and feedback loops, including Planning–Instructor debate, introduce computational latency incompatible with millisecond-level wireless control tasks.This limitation motivates architectural separation between deliberative reasoning and time-critical control.
  • B. Balancing reasoning capabilities with real-time latency: A hierarchical control architecture should bridge slow large-model inference with strict physical-layer timing requirements.The proposed division assigns intent understanding and strategic planning to a high-level control plane.
  • B. Balancing reasoning capabilities with real-time latency: Time-critical execution should be delegated to lightweight specialized small models or conventional controllers to prevent agent reasoning from bottlenecking signal processing.This arrangement supports network stability despite alternating control between components with different response times.

C. Coordination scalability in distributed architectures · D. Long-term memory and lifelong adaptation · VI. CONCLUSIONS

The paper identifies coordination overhead and missing long-term memory as scalability limitations, then concludes that ComAgent’s multi-agent cognitive cycle enables autonomous wireless-network design while mitigating monolithic LLM weaknesses.

  • C. Coordination scalability in distributed architectures: As networks grow, exchanging lengthy prompts and feedback becomes a critical coordination bottleneck.The framework must address communication costs when supporting large and complex networks.
  • C. Coordination scalability in distributed architectures: Centralized orchestration creates a single-point-of-failure bottleneck and high signaling overhead, while full decentralization risks inconsistency.The passage frames both architectural extremes as problematic for scalable coordination.
  • C. Coordination scalability in distributed architectures: Future systems should use adaptive agent topologies that dynamically form task-based subgroups.This is proposed as a response to coordination challenges in large and complex networks.
  • D. Long-term memory and lifelong adaptation: Current case studies retrieve literature anew for each task and lack retained experience across sessions.This causes inefficient duplication of effort for recurring optimization problems in dynamic wireless environments.
  • D. Long-term memory and lifelong adaptation: Future research should integrate long-term experiential memory to archive valid solution templates and failure cases.This memory should be distinct from temporary system memory and support adaptation across sessions.
  • VI. CONCLUSIONS: ComAgent autonomously navigates the complex design space of emerging wireless networks through a multi-LLM agentic AI framework.The framework addresses the complexity limitations of traditional manual optimization.
  • VI. CONCLUSIONS: Specialized Literature, Planning, Coding, and Scoring Agents are coordinated through a recursive Perception-Planning-Action-Reflection cycle.This orchestration mitigates hallucination and reasoning deficits inherent in monolithic LLMs.
  • VI. CONCLUSIONS: Extensive case studies, including MIMO SWIPT beamforming and a generic optimization test set, validated ComAgent’s automated end-to-end workflow.The supplied conclusion identifies these case studies as validation settings.
Loading 2601.19607v1…