Comagent Achieves 6G Network Optimisation Via Multi-LLM Agentic AI

Scientists are tackling the challenge of translating complex network requirements into workable solutions for future 6G systems. Haoyun Li, Ming Xiao (KTH Royal Institute of Technology), and Kezhi Wang (Brunel University London) , alongside Schober et al , introduce ComAgent, a novel multi-LLM framework designed to automate the process of wireless network design and optimisation. This research is significant because it moves beyond the limitations of single Large Language Models by creating a closed-loop system that can independently search for information, write code, and verify its own results. ComAgent effectively bridges the gap between high-level intentions and practical implementation, achieving performance comparable to human experts and promising a revolution in how we build and manage next-generation wireless networks.

This breakthrough research addresses a fundamental bottleneck in modern wireless network development: the time-intensive and error-prone process of translating abstract goals into mathematically sound and executable simulations. ComAgent distinguishes itself from monolithic LLM approaches by incorporating explicit domain grounding, robust constraint awareness, and a crucial execution-based verification process. The framework iteratively decomposes complex problems, self-corrects errors, and generates solver-ready formulations, ensuring both accuracy and feasibility, a significant advancement over existing methods.

Experiments show ComAgent achieves expert-comparable performance in challenging beamforming optimisation tasks, demonstrating its ability to design algorithms that rival those created by human specialists. Evaluations demonstrate the efficacy of the framework through a non-trivial beamforming optimisation case study, where the system autonomously perceives the problem, designs an algorithm, and generates high-performing solutions. This work opens exciting possibilities for automating network design and operation, reducing reliance on manual intervention and accelerating the deployment of next-generation wireless technologies. Furthermore, the research establishes a new paradigm for network management, shifting from rigid mathematical optimisation to a flexible, semantic-aware control system. By leveraging the superior generalization and reasoning capabilities of LLMs, ComAgent bypasses the need for explicit mathematical modelling, streamlining complex network orchestration. The framework’s ability to adapt to new tasks with zero- or few-shot learning represents a significant step towards truly intelligent and autonomous wireless networks, capable of dynamically responding to evolving service demands and heterogeneous infrastructures.

ComAgent, Automated 6G Network Modelling via LLMs

This work addresses the bottleneck of manually creating mathematical models for complex 6G network optimisation, employing a closed-loop Perception-Planning-Action-Reflection cycle to achieve autonomous operation. The team engineered a system coordinating specialised agents, Literature, Planning, Coding, and Scoring, to iteratively decompose problems and self-correct errors, bridging the gap between user requests and executable code. Initially, the Literature Agent undertakes targeted searches to retrieve relevant domain knowledge, grounding the subsequent steps in established research and standards. Following this, the Planning Agent decomposes the initial high-level intent into a series of structured sub-tasks, defining a clear solution strategy and outlining the necessary mathematical formulations.

Crucially, the Coding Agent then translates these formulations into executable simulation code, utilising specific programming languages and optimisation libraries to create a functional model. This agent leverages retrieved knowledge to ensure the code adheres to physical constraints and network standards, minimising the risk of hallucinations or constraint violations. The generated code is subsequently executed within a simulation environment, and the resulting performance metrics are fed back to the Scoring Agent. This agent evaluates the solution against predefined criteria, such as data rate, latency, and energy efficiency, providing a quantitative score that reflects its effectiveness.

The system then enters a recursive loop, utilising the score to refine the planning and coding stages, iteratively improving the solution until it meets the desired performance targets. Experiments employed a non-trivial MIMO SWIPT beamforming case study to demonstrate ComAgent’s capabilities. Evaluations revealed that ComAgent achieves expert-comparable performance in complex beamforming optimisation, demonstrating its ability to generate high-quality solutions autonomously. Furthermore, the study highlights that ComAgent outperforms monolithic LLMs across diverse wireless tasks, showcasing its potential for automating design in emerging wireless networks and significantly reducing reliance on human domain expertise. This innovative approach enables a paradigm shift from rigid mathematical optimisation to flexible, semantic-aware control, streamlining complex network orchestration and facilitating end-to-end network management.

ComAgent automates 6G network design via LLMs

The research addresses a critical bottleneck in current network design, which relies heavily on manual mathematical modelling and algorithm creation. ComAgent employs a Perception-Planning-Action-Reflection cycle, coordinating specialized LLMs for tasks including literature search, coding, and scoring, to generate reproducible simulations autonomously. The team measured the system’s ability to independently perceive the problem, design an appropriate algorithm, and generate solutions matching expert baselines. The breakthrough delivers a system capable of creating complete, reproducible simulation pipelines while simultaneously identifying and correcting logical and feasibility errors.

Scientists recorded that ComAgent’s iterative decomposition and self-correction mechanisms significantly reduce the need for human intervention in the optimization process. Measurements confirm that the framework can handle complex, high-dimensional problems with mixed discrete, continuous decisions and intricate coupling constraints, a common challenge in modern wireless network design. Tests prove that ComAgent’s closed-loop architecture enables continuous improvement and adaptation to new scenarios without requiring extensive retraining or expert knowledge. This work paves the way for more flexible, scalable, and intelligent wireless infrastructure capable of meeting the demands of future applications like extended reality and autonomous systems.

ComAgent outperforms LLMs in wireless optimisation, achieving superior

This innovative system employs a Perception-Planning-Action-Reflection cycle, utilising specialised agents for tasks such as literature review, coding, and performance scoring, to translate user intentions into workable solutions. This demonstrates its potential to significantly streamline the design process for future 6G networks, automating workflows from mathematical modelling to simulation implementation. The authors acknowledge a limitation in the system’s current reliance on temporary memory, suggesting future work could integrate long-term experiential memory to archive successful strategies and avoid repeating design loops for familiar scenarios. This research establishes a foundational multi-LLM agentic AI architecture for intent-driven network management, potentially paving the way for self-evolving and zero-touch 6G network systems. By mitigating the common issues of hallucination and reasoning deficits found in single LLMs, ComAgent offers a robust and adaptable solution for the increasing complexity of modern wireless network design. Further development incorporating long-term memory could enhance its robustness and adaptability over time, moving beyond reactive problem-solving towards proactive optimisation.