Multi-LLM Agentic AI at the Edge: Joint Agent Scheduling and Resource Allocation
Organizations: Department of Computer Science, Brunel University of London, UB8 3PH, UK. · Key Laboratory of Broadband Wireless Communication and Sensor Network Technology, Nanjing University of Posts and Telecommunications, Nanjing 210003, China.
Abstract
In this paper, an edge-enabled multi-large language model (LLM) agentic AI architecture is proposed, where one supervisor AI agent and multiple worker AI agents cooperatively execute users' requests at the edge. The supervisor AI agent decouples the user request into a series of sub-tasks and dynamically dispatches them to LLM-equipped worker AI agents via two agentic workflows, i.e., parallel and serial workflows, with a two-level retry mechanism to enhance execution reliability. To prolong the service duration of worker AI agents, the energy minimization problems are formulated by jointly considering AI agent selection and resource allocation, subject to task execution latency and quality of service (QoS) constraints. A resource allocation scheme based on alternating optimization (AO) is proposed. Specifically, for the parallel workflow, the scheme decouples the original problem into bandwidth allocation, retrieval selection and AI agent selection sub-problems, which are iteratively solved by the proposed successive convex approximation (SCA), Lagrangian relaxation (LR), and Gale-Shapley (GS) or LLM-based algorithms, respectively. For the serial workflow, the problem is similarly decoupled into service time allocation, retrieval selection and AI agent selection sub-problems; the optimal service time allocation is obtained via Karush-Kuhn-Tucker (KKT) conditions, while the latter two sub-problems are addressed by extending the proposed algorithms to the serial scenario. Simulation results show that the proposed resource allocation scheme can reduce system energy consumption compared with other benchmarks, while experiments on a multi-LLM agentic AI system based on Qwen and Llama demonstrate improved task execution success rates.
Figures & tables
| Notation | Value | Notation | Value |
|---|---|---|---|
| 30 | 1 W | ||
| W | -90 dB | ||
| -85 dB | |||
| MHz | -174 dBm/Hz | ||
| {28,32,36} |
| LLM optimizer | Mean ( J) | Variance ( J 2 ) |
|---|---|---|
| ChatGPT-5.4 | 7.776 | 5.324 |
| ChatGPT-Turbo | 11.48 | 9.862 |
| Model name | Success rate | Tool call | Retry | |
|---|---|---|---|---|
| Serial | Parallel | |||
| Qwen2.5-1.5B-Instruct | 3.37% | 4.23% | 37.42% | 67.80% |
| Qwen2.5-3B-Instruct | 7.87% | 5.76% | 60.31% | 42.10% |
| Qwen2.5-7B-Instruct | 12.36% | 8.27% | 66.00% | 35.27% |
| Qwen2.5-Math-1.5B-Instruct | 0.00% | 0.00% | 0.90% | 99.33% |
| Qwen2.5-Math-7B-Instruct | 0.22% | 0.41% | 1.81% | 98.62% |
| Scheme | Task execution success rate | Tool call success rate | ||||
|---|---|---|---|---|---|---|
| Group 1 | Group 2 | Group 1 | Group 2 | |||
| Serial | Parallel | Serial | Parallel | |||
| RandAS | 6.05% | 6.44% | 9.98% | 9.69% | 34.14% | 53.31% |
| DRL-PPO | 8.49% | 7.65% | 21.93% | 23.74% | 69.56% | 73.96% |
| AO-GS | 8.28% | 7.86% | 20.51% | 22.53% | 69.71% | 71.28% |
| AO-LLM | 9.38% | 8.86% | 23.29% | 25.45% | 71.83% | 79.14% |
| Schemes | Without retry | ChatGPT-5.4 | Qwen-plus | Llama-4-Maverick |
|---|---|---|---|---|
| AO-GS | 16.24% | 19.82% | 21.61% | 23.51% |
| AO-LLM | 17.55% | 21.24% | 24.92% | 28.93% |
| Yuanyan Song received the B.S. and M.S. degree from Wuhan University of Technology, Wuhan, China, in 2022 and 2025, respectively. He is currently pursuing the Ph.D. degree with Brunel University of London, London, UK. His research interest is large language models, agentic AI and wireless networks. |
| Kezhi Wang (Senior Member, IEEE) received the PhD degree from the University of Warwick, U.K. He is currently a Professor in the Department of Computer Science at Brunel University of London, UK. His research interests include wireless communications, edge computing, and machine learning. |
| Xinmian Xu is currently pursuing the Ph.D. degree with Nanjing University of Posts and Telecommunications, Nanjing, China.,His research interests include multimodal communications, semantic communication, resource allocation, and signal processing in wireless networks. |
| Yongmin Li (Senior Member, IEEE) received the BEng and MEng degrees from Tsinghua University, China, and the PhD degree from Queen Mary, University of London, UK. He previously worked as a research scientist at British Telecom Laboratories. He is currently a Professor in the Department of Computer Science, Brunel University London. His research interests include artificial intelligence, natural language processing, data science, control engineering, image processing, healthcare technologies, and human computer interaction. He has been involved in numerous research projects funded by EPSRC, BBSRC, MRC, Innovate UK, NSFC and CSC, and has served as an editor for several leading academic journals and as a panel member for research councils and committees. He is a multiple award winner, with awards including the Innovate UK Knowledge Transfer Partnerships Outstanding Award 2024, the MICCAI RETOUCH Grand Challenge 2023, and the MVA Most Influential Paper over the Decade Award 2019. He is a Senior Fellow of the Higher Education Academy. |
Multi-Agent LLM Systems
Multi-agent large language model systems coordinate multiple AI agents to solve tasks through communication, specialization, shared tools, and memory.
See all 1,108 papers