Paper deep dive
LLM-Based Multi-Agent Systems over Wireless Networks: A Joint Agent--Network Design Perspective
Chao Hu, Yuan Guo, Guanlin Wu, Yueling Che, Han Hu, Jie Xu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 9/30/2026, 6:27:34 AM
Summary
This paper proposes a joint agent-network design perspective for Large Language Model (LLM)-based Multi-Agent Systems (MASs) operating over wireless networks. It addresses the coupling between multi-agent reasoning dependencies and wireless network constraints (connectivity, edge resources). The authors identify six technical challenges: metric misalignment, message redundancy, state inconsistency, topology mismatch, resource limitation, and trust discontinuity. To resolve these, they propose joint designs for interaction scheduling/resource allocation, message selection/transmission, and topology/workload allocation, supported by network-verified provenance. A V2X case study demonstrates that this joint approach outperforms separate agent-only or wireless-only designs in task completion under constraints.
Entities (12)
Relation Signals (9)
LLM-based multi-agent systems â operatesover â Wireless Networks
confidence 95% · LLM-based Multi-Agent Systems (MASs) over wireless networks... their reasoning and execution are increasingly distributed across wireless edge nodes.
Joint Agent-Network Design â outperforms â Separate Designs
confidence 94% · An illustrative vehicle-to-everything (V2X) case study shows that jointly adapting agent-side interaction decisions and network operations... outperforming the conventional agent-only and wireless-only separate designs.
Joint Agent-Network Design â addresses â Message Redundancy
confidence 92% · Specifically, we present... the message selection-transmission co-design...
Joint Agent-Network Design â addresses â State Inconsistency
confidence 92% · Specifically, we present... the message selection-transmission co-design...
Joint Agent-Network Design â addresses â Resource Limitation
confidence 92% · Specifically, we present... the joint agentânetwork topology design and workloadâresource allocation.
Joint Agent-Network Design â addresses â Topology Mismatch
confidence 92% · Specifically, we present... the joint agentânetwork topology design and workloadâresource allocation.
Joint Agent-Network Design â addresses â Metric Misalignment
confidence 92% · To address these challenges, this article develops a novel joint agentânetwork design perspective... Specifically, we present the joint design of agentâinteraction scheduling and resource allocation...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:As large language models (LLMs) evolve from standalone models into collaborative agents embedded in physical systems, their reasoning and execution are increasingly distributed across wireless edge nodes. In this setting, wireless networks are experiencing a paradigm shift from only providing data connectivity to supporting the multi-agent reasoning workflow itself. The task performance of such network-constrained LLM-based multi-agent systems (MASs) is jointly affected by the multi-agent reasoning dependencies as well as the underlying network connectivity and edge resources. This coupling gives rise to various technical challenges, including the metric misalignment and message redundancy, state inconsistency and topology mismatch, as well as resource limitation and trust discontinuity. To address these challenges, this article develops a novel joint agent--network design perspective that coordinates decisions on both sides of the system. Specifically, we present the joint design of agent--interaction scheduling and resource allocation, the message selection-transmission co-design, as well as the joint agent--network topology design and workload--resource allocation. Furthermore, we consider the network-verified provenance that is linked with agent-side information-flow control to constrain how received information affects subsequent operations. An illustrative vehicle-to-everything (V2X) case study shows that jointly adapting agent-side interaction decisions and network operations improves task completion under communication and edge-resource constraints, outperforming the conventional agent-only and wireless-only separate designs.
Tags
Links
- Source: https://arxiv.org/abs/2609.37094v1
- Canonical: https://arxiv.org/abs/2609.37094v1
Trouble viewing inline? Open PDF directly â
Full Text
38,349 characters extracted from source content.
Expand or collapse full text
1 LLM-Based Multi-Agent Systems over Wireless Networks: A Joint AgentâNetwork Design Perspective Chao Hu, Graduate Student Member, IEEE, Yuan Guo, Member, IEEE, Guanlin Wu, Graduate Student Member, IEEE, Yueling Che, Member, IEEE, Han Hu, Senior Member, IEEE, and Jie Xu, Fellow, IEEE AbstractâAs large language models (LLMs) evolve from stan- dalone models into collaborative agents embedded in physical sys- tems, their reasoning and execution are increasingly distributed across wireless edge nodes. In this setting, wireless networks are experiencing a paradigm shift from only providing data connectivity to supporting the multi-agent reasoning workflow itself. The task performance of such network-constrained LLM- based multi-agent systems (MASs) is jointly affected by the multi-agent reasoning dependencies as well as the underlying network connectivity and edge resources. This coupling gives rise to various technical challenges, including the metric misalignment and message redundancy, state inconsistency and topology mis- match, as well as resource limitation and trust discontinuity. To address these challenges, this article develops a novel joint agentâ network design perspective that coordinates decisions on both sides of the system. Specifically, we present the joint design of agentâinteraction scheduling and resource allocation, the message selection-transmission co-design, as well as the joint agentâ network topology design and workloadâresource allocation. Fur- thermore, we consider the network-verified provenance that is linked with agent-side information-flow control to constrain how received information affects subsequent operations. An illustrative vehicle-to-everything (V2X) case study shows that jointly adapting agent-side interaction decisions and network operations improves task completion under communication and edge-resource constraints, outperforming the conventional agent- only and wireless-only separate designs. Index TermsâLLM agents, multi-agent systems, wireless net- works, joint agentânetwork design, 6G. I. INTRODUCTION Large language models (LLMs) are extending artificial intelligence (AI) from passive inference toward autonomous agents. These agents can reason over context, invoke external tools, and act toward specific objectives [1], [2]. As tasks become more complex with heterogeneous information and reasoning requirements, a single agent may become insuffi- cient. This limitation motivates LLM-based multi-agent sys- tems (MASs), where the reasoning process is distributed across multiple agents with complementary roles [3], [15]. C. Hu is with the College of Computer Science and Software Engineering (CSSE), Shenzhen University, Shenzhen 518060, China, and also with the School of Science and Engineering (SSE), the Shenzhen Future Network of Intelligence Institute (FNii), and the Guangdong Provincial Key Laboratory of Future Networks of Intelligence, The Chinese University of Hong Kong (Shenzhen), Shenzhen 518172, China (e-mail: chaohu@link.cuhk.edu.cn). Y. Guo, G. Wu, and J. Xu are with the SSE, the FNii, and the Guangdong Provincial Key Laboratory of Future Networks of Intelligence, The Chinese University of Hong Kong (Shenzhen), Shenzhen 518172, China (e-mail: guoyuan@cuhk.edu.cn; guanlinwu1@link.cuhk.edu.cn; xujie@cuhk.edu.cn). Y. Che is with the CSSE, Shenzhen University, Shenzhen 518060, China (e-mail: yuelingche@szu.edu.cn). H. Hu is with the School of Information and Electronics, Beijing Institute of Technology, Beijing 100081, China (e-mail: hhu@bit.edu.cn). Y. Guo and Y. Che are the corresponding authors. On the other hand, these MASs are increasingly deployed in distributed physical systems, where agents rely on observations collected at different locations and must update their decisions under time-critical requirements. In this case, conventional cloud-centric architectures, in which the reasoning needs to be implemented at far-apart cloud nodes with increased com- munication and execution delay, are poorly matched to multi- agent reasoning. By contrast, edge computing is becoming particularly appealing, which places reasoning at edge nodes close to the observation sources, allowing newly acquired observations to enter the reasoning process quickly [4]. In such scenarios, the reasoning process must account for wireless net- work connectivity because inter-agent information exchange is carried over wireless links. Poor connectivity can delay or interrupt the information required for subsequent reasoning steps, thus degrading the performance of task completion [5]. However, conventional agent systems and wireless com- munication networks are generally designed separately. On the one hand, an LLM-based MAS organizes collaboration through an interaction structure that determines which agents exchange information during reasoning. At the initial stage, this structure was prescribed in advance through fixed roles and predefined interaction patterns [2], [3]. As collaborative tasks become more diverse, such fixed structures cannot adapt to the current reasoning demand. Recent work has therefore made the interaction structure adaptive. For instance, Agent- Prune removes exchanges with low task contribution [9], while G-Designer generates task-dependent interaction topologies from the current collaboration context [10]. Nonetheless, this formation process still largely ignores the network architecture and connectivity through which the selected interactions must be realized. On the other hand, conventional wireless network design typically assumes a given network topology and commu- nication demand, and then allocates wireless resources to support the corresponding transmissions. Conventionally, wire- less design focused on reliable bit delivery under channel and resource constraints. Recently, task-oriented and prag- matic communication introduced task utility into transmission design [6], but these approaches mainly optimize how a given interaction is delivered at the link level. They do not determine how inter-agent collaboration should be formed in an MAS. More recently, distributed LLM inference has been designed coupling with wireless transmission with edge com- putation [7], while agentic communication has begun to relate communication decisions to the reasoning process [8]. Recent studies have also considered multi-agent coordination over 6G networks [3], [4]. Nevertheless, how agent interactions arXiv:2609.37094v1 [cs.MA] 29 Sep 2026 2 Fig. 1. From single-agent operation to multi-agent collaboration over wireless networks. Task planning forms inter-agent collaboration, while its runtime execution and adaptation depend on wireless connectivity and edge resources. should be formed according to both reasoning requirements and network conditions remains not well understood. Motivated by this gap, this article studies joint agent and network design for LLM-based MASs over wireless networks. We identify six technical challenges, including metric mis- alignment and message redundancy, state inconsistency and topology mismatch, as well as resource limitation and trust discontinuity. We then develop various joint design approaches that use task utility to connect agent-side decisions with network-side communication and resource allocation. An il- lustrative vehicle-to-everything (V2X) case study compares the joint design with two conventional separate designs, where the agent-only design adapts the agent-side information decision with fixed network resources, while the wireless-only design allocates network resources for a fixed agent-side decision. The results show that the joint design improves task comple- tion under communication and edge-resource constraints and outperforms both separate designs. I. NETWORK-CONSTRAINED LLM-BASED MASS This section establishes the system foundation for LLM- based MASs over wireless networks. We first describe an indi- vidual LLM agent and then extend to multi-agent collaboration over wireless network edge 1 . We then discuss representative applications and the resulting technical challenges. 1 The discussion also applies to agentic systems built on other multimodal foundation models, including multimodal large language models (MLLMs), vision-language models (VLMs), as well as vision-language-action (VLA) models. A detailed treatment of multimodal perception and action is beyond the scope of this article. A. System Basics An LLM-based agent is an autonomous decision entity that uses an LLM as its reasoning core, with memory and external tools supporting its operation, as illustrated in the left panel of Fig. 1. At each reasoning step, the agent forms a context, defined as the information currently available to the LLM from its input and retained memory. The LLM processes this context to determine the next action, and the agent then updates the context based on newly obtained information. This process continues as the task progresses. An LLM-based MAS consists of multiple LLM agents that cooperate on a shared task. One important issue faced by such systems is reasoning dependency, i.e., one agent generally needs to acquire the information or a processing result from another before the corresponding reasoning step can continue. Here, the message exchanged to satisfy such a dependency is referred to as an interaction. When the agents are hosted on different network nodes, these messages are exchanged over wireless communication links. Wireless connectivity can therefore affect when the required information reaches the receiving agent and, consequently, whether the dependent reasoning step can proceed. Since wireless links may vary during task execution, changes in message delivery can further affect the subsequent multi-agent collaboration. LLM agents can be hosted at different locations across the network (e.g., at end devices, edge nodes, and cloud servers). Agents hosted on end devices remain close to local observations and physical actions, but have limited computing capacity. Agents hosted at nearby edge nodes can access addi- tional edge computing resources while remaining close to the devices they support. Remote cloud servers can accommodate larger models or longer context, but the longer communi- 3 TABLE I JOINT DESIGN SPACE FOR NETWORK-CONSTRAINED LLM-BASED MASS. ChallengeAgent SideNetwork/Edge SideCouplingJoint Decision Metric misalignmentInteraction priority [11]Resource allocationInteraction creditâ service feasibility Interactionâresource scheduling Message redundancy and state inconsistency Message selection/compression [9] Delivery fidelity/timeliness [6], [8] Receiver-context-conditioned decision distortion Message selection and adaptive delivery Topology mismatchLogical edge/endpoint selection [9], [10], [15] Association/path selection [5], [8] Interaction creditâ support cost Joint edgeânetwork mapping Resource limitationReasoning budget/workload Communication/edge- resource allocation [4], [7] Workload footprintâ resource feedback Workloadâresource allocation Trust discontinuityInformation-flow control [13] Provenance verification/admission Provenanceâ trust requirement Provenance-aware information flow cation path makes cloud execution less suitable for time- critical reasoning. Agents hosted across these tiers can also collaborate during task execution, allowing latency-sensitive reasoning to remain close to the information source while more resource-demanding reasoning is supported by edge or cloud agents. Such cross-tier collaboration therefore creates a trade- off between communication distance and available execution resources, which affects whether a reasoning dependency can be completed within the task deadline. The right panel of Fig. 1 illustrates how multiple agents col- laborate during task execution. We use the term collaboration workflow to denote the runtime process that converts a shared task into reasoning dependencies and maps these dependencies to multi-agent interactions. Newly obtained information may change the unresolved dependencies during execution, causing the interaction structure to be updated. The collaboration workflow therefore determines how inter-agent collaboration evolves as the task progresses. Table I summarizes the coupling between this workflow and network operation. A reasoning dependency does not necessarily correspond to a unique interaction because dif- ferent agents may provide information that satisfies the same task requirement. The MASs determine which interaction is useful for the current reasoning process, while the network determines whether that interaction can be supported un- der the available connectivity and resources. Task execution is therefore constrained by both the multi-agent interaction workflow and the underlying network, which raises the joint agent and network design challenges, as will be discussed in Section I-C. B. Representative Applications Fig. 2 summarizes representative applications of LLM-based MASs over wireless networks. Vehicle-to-Everything (V2X) Coordination: In coopera- tive perception, an ego vehicle may require observations from nearby nodes to identify an occluded pedestrian or vehicle. An LLM agent hosted on a nearby vehicle or roadside unit (RSU) can provide the missing information, while the wireless network determines whether the message can reach the ego vehicle within the required decision time. Uncrewed Aerial Vehicle (UAV) Swarm: In a search or monitoring mission, UAV agents can divide the sensing area and exchange local observations during execution. When a UAV completes its assigned region or loses a reliable link, the remaining agents may need to redistribute the sensing task. The resulting collaboration therefore depends on both the evolving mission and the communication links among the UAVs. Networked Robotics: In networked robotic systems, agents may be hosted on manipulators, mobile robots, or local con- trollers that perform different parts of a shared task. A robot may require an observation or execution result from another robot before its next action can proceed. LLM-based MASs can adapt the subsequent task assignment according to these intermediate results, while wireless connectivity determines whether the required information is available in time. Network Management: In this scenario, the wireless net- work serves as both the communication infrastructure and the object being controlled. Distributed agents can diagnose a network problem and modify its configuration. The resulting network conditions then affect the communication available for subsequent agent collaboration. On-Demand Agent Coalition: In emerging tasks such as disaster response or emergency search, the participating agents may not be known in advance. An LLM-based MAS can identify available agents according to the current task and form a temporary coalition using agents hosted on mobile or infrastructure nodes. The wireless network determines which of these candidate agents can participate under the current connectivity and resource conditions. C. Technical Challenges Network-constrained LLM-based MASs face various tech- nical challenges when multi-agent reasoning is executed over wireless networks. As summarized in Fig. 2, these challenges arise from task evaluation and information exchange, as well as from the network realization and trusted execution of agent interactions. 2 1) Metric Misalignment: Wireless systems typically charac- terize transmission performance using quality-of-service (QoS) 2 The six challenges span task evaluation and information transfer, in- teraction realization and resource-constrained execution, as well as trusted information use. Message redundancy and state inconsistency are addressed together because both concern whether task-relevant information reaching the receiver remains useful for its subsequent reasoning. 4 Fig. 2. Key challenges and representative applications of LLM-based MASs over wireless networks. metrics such as throughput and delay. These metrics describe how effectively information is delivered, but not how much a particular interaction contributes to the multi-agent task. This contribution is difficult to measure, as its effect may become observable only after subsequent reasoning has unfolded, whereas wireless scheduling must prioritize transmissions be- fore the task outcome is known. The network therefore lacks a directly measurable interaction-level metric that connects immediate transmission decisions to downstream task utility. 2) Message Redundancy: Message redundancy occurs when agents exchange observations or reasoning results that add little new information to the receiver [9]. Over wireless links, such redundant message transmission consumes communi- cation resources and may delay more useful interactions. Reducing such redundancy is essential but difficult because content similarity does not determine a messageâs contribution to the receiverâs next decision. A message similar to the existing context may provide a missing fact, while a different message may have little effect on the current reasoning. Its usefulness must therefore be evaluated with respect to the receiverâs current context and subsequent decision. 3) State Inconsistency: The state of an agent refers to the context used for its next reasoning step. Here, context denotes the information available to the LLM for reasoning, while state emphasizes the information whose consistency affects coordi- nated execution. Different agents may retain inconsistent states when a shared update is delayed or missed during wireless transmission. Nevertheless, reliable delivery alone does not remove this problem because an update may arrive after the receiving agent has already reasoned from an outdated state. Maintaining the freshness of message updates is therefore important but difficult, as the required freshness depends on how an update affects the subsequent reasoning step. Packet delivery status alone does not reveal this requirement. 4) Topology Mismatch: An MAS forms logical interactions according to reasoning dependencies, while the wireless net- work provides physical connectivity between the nodes hosting the agents. A topology mismatch occurs when a task-relevant agent is difficult to reach through the available links, whereas a well-connected agent may provide little useful informa- tion [4], [10]. Therefore, interaction selection is important but challenging, as it depends on both task relevance and physical reachability. Neither the agent interaction graph nor the wireless connectivity alone determines which interaction should be realized. 5) Resource Limitation: Distributed LLM agents rely on local sensing capabilities, while communication and computa- tion resources at edge nodes are limited [4], [7]. The resulting workload is not fully known before execution because received context can change the amount of subsequent reasoning. Intermediate reasoning results may also trigger new inter- agent exchanges. Consequently, communication and execution demands can evolve during the collaboration itself. Allocating communication and computation resources is therefore essen- tial but difficult because the demand being served is partly generated by the same reasoning process whose execution depends on those resources. 6) Trust Discontinuity: Wireless security mechanisms can authenticate a sender and verify the integrity of a delivered message, but these guarantees do not determine how the received information should be allowed to influence subse- quent agent reasoning and actions. The corresponding design issue is therefore to preserve network-verified provenance and use it to constrain downstream information use. This is important because an authenticated message may still contain a prompt injection or other malicious instruction that alters the receiverâs behavior [12]. It is also difficult because provenance established during delivery may no longer remain explicit after the information is incorporated into the agentâs context or memory. Network-side verification cannot determine the later semantic influence of the message, while agent-side control may lack the provenance needed to enforce source- dependent policies. Secure multi-agent reasoning therefore requires network-established provenance to remain available as information propagates through the agentâs reasoning and execution process [13]. 5 I. JOINT AGENT AND NETWORK DESIGN To resolve the above challenges, we pursue joint agent and network design based on task requirements and network conditions. Task utility serves as the primary performance objective, while trust requirements impose constraints on how information can influence agent operations. The other designs identify the coupled agent-side and network-side decisions under these objectives and constraints. A. Joint Interaction Scheduling and Resource Allocation Metric misalignment requires interaction scheduling and network resource allocation to be considered jointly. An MAS may select an interaction with a high contribution to task completion even when the communication link supporting that interaction cannot deliver its message before the task dead- line. Conversely, a network-side scheduler that considers only queue and channel conditions may allocate more resources to a link carrying an interaction with little contribution to task completion. Hence, interaction activation and resource allocation should account for both task contribution and link feasibility. Credit assignment provides the task-side metric needed for this decision. In MASs, credit denotes the contribution of an agent or action to a shared task outcome [11]. We extend this notion to an inter-agent exchange and define its interaction credit as the estimated marginal contribution of that exchange to task utility under a given delivery condi- tion. Offline evaluations can approximate this contribution by comparing task outcomes when the interaction is available with those obtained when it is withheld or delayed under otherwise matched settings. A runtime predictor can then estimate the interaction credit from the current task context and expected delivery condition. The MASs use the predicted credit to prioritize candidate interactions, while the network provides the corresponding link feasibility or service cost. These quantities jointly determine which interactions are acti- vated and how resources are allocated. A modified max-weight policy provides one possible implementation by incorporating interaction credit into the conventional service weight. B. Joint Message Selection and Transmission Message redundancy and state inconsistency require the system to determine what information should be transmitted and how it should be delivered. Agent-side message selection alone may retain information that is useful in isolation but redundant to the receiverâs current context or ineffective when delivered too late. Network-side transmission, in contrast, treats the selected content as given and does not know which part will affect the receiverâs next reasoning step. A joint design therefore requires a common receiver-side criterion for message selection and transmission. Communication-efficient methods provide the agent-side mechanism for selecting message content [9], while task- oriented joint sourceâchannel coding (JSCC) provides a network-side mechanism for adapting transmission to an end-task loss [6], [8]. These mechanisms can be coupled through a receiver-context-conditioned decision distortion, which measures how removing, delaying, or distorting infor- mation changes the receiverâs subsequent decision. The MASs use this metric to retain or compress information according to its decision impact. The transmission scheme uses the same metric to determine the required delivery fidelity. For a time-sensitive state update, delayed delivery produces a larger penalty when the update no longer supports the receiverâs next reasoning step. C. Joint Agent and Network Topology Design Topology mismatch requires the logical interaction graph and its network realization to be configured together. Adaptive MASs can construct or prune interaction graphs according to reasoning dependencies, with the interaction credit introduced above representing the task value of each candidate edge [9], [10]. Wireless association and routing instead determine how the nodes hosting the selected agents can be connected under current channel and resource conditions. If these decisions are separated, a high-value logical edge may be difficult to realize, while a network-efficient connection may correspond to an interaction with little task value. Task relevance and network feasibility should therefore enter the same topology decision. For each unresolved reasoning dependency, the MASs iden- tify candidate agents and assigns an interaction credit to the corresponding logical edges. The network then evaluates fea- sible associations or communication paths for each candidate and computes a support cost, defined as the communication resources required to satisfy its delivery constraints. The logi- cal edge and its network realization are subsequently selected according to interaction credit and support cost under the available connectivity and resources. Because both choices are discrete, the resulting edgeânetwork mapping forms a combinatorial optimization problem. Existing task-dependent graph construction and wireless association or routing provide the two building blocks for this joint topology design. D. Joint Workload Control and Resource Allocation Resource limitation couples the workload generated by the MASs with the resources available to execute it. An agent-side controller may select a more demanding reasoning configura- tion for greater expected task gain even when the resulting workload cannot be completed within the available resources and task deadline. Conversely, conventional resource allocation treats the submitted workload as fixed or uncontrollable and only determines how the available resources should serve it. Workload control and resource allocation should therefore be determined together according to expected task gain and execution feasibility. Here, agentic workload denotes the communication and execution demand induced by a reasoning configuration. Reasoning-budget control provides one agent-side mechanism for changing this workload, where the reasoning budget bounds the extent of reasoning performed by the agents. Distributed inference and computation offloading provide network-side mechanisms for allocating communication and edge resources [4], [7]. A workload profiler can connect the 6 two by predicting the resource footprint induced by a reason- ing configuration without executing the complete reasoning process. The reasoning configuration and resource allocation can then be jointly optimized to maximize expected task utility under the task deadline and the available communication and computing resources. For larger design spaces, marginal resource prices derived from active constraints can provide compact feedback for adapting the reasoning budget as net- work and edge conditions change. E. Joint Provenance Verification and Information-Flow Con- trol Trust discontinuity requires network-side provenance ver- ification to remain connected with agent-side control over how received information is used. Network authentication and integrity verification establish the source and integrity of a delivered message, while information-flow control (IFC) restricts how that information can propagate through reasoning and execution [13]. If these mechanisms operate separately, an authenticated message may still contain malicious in- structions [12], and the agent may later lack the verified provenance needed to enforce source-dependent policies. Joint design therefore needs to preserve verified provenance after delivery and enforce it when the information influences agent operations. One implementation is to bind the verified source and integrity attributes of an incoming message to its content as a provenance label. Agent-side IFC propagates this label when the information enters context or memory and when it contributes to derived results. Before a protected operation, the agent checks whether the provenance associated with the influencing information satisfies the corresponding policy. The coupling can also operate in the reverse direction. If a reasoning dependency requires information from a specified trust class, this requirement can constrain source selection or message admission. Network verification and agent-side IFC therefore connect the origin of received information with the permissions governing its downstream influence. IV. ILLUSTRATIVE CASE STUDY: V2X MULTI-AGENT COLLABORATION To illustrate joint agent and network design, we consider a concrete V2X task in an unprotected left-turn scenario with an occluded pedestrian, as shown in Fig. 3. The ego vehicle V 1 must decide whether it is safe to enter the conflict zone, while a nearby truck blocks its view of the pedestrian. Vehicle V 2 has a complementary view of the occluded region, whereas V 3 provides additional traffic information [14]. Each vehicle hosts an LLM-based agent, while the RSU serves as a base station (BS)-like wireless access point and hosts the coordinator agent. All vehicles communicate with the RSU through vehicle-to-infrastructure links; no direct vehicle-to- vehicle (V2V) or device-to-device (D2D) link is considered. This physical topology remains fixed for all evaluated designs, so the case study focuses on interaction, message, and resource decisions rather than topology optimization. Vehicle-side ex- ecution resources are fixed, while the uplink bandwidth and Fig. 3. Illustration of case scenario and collaboration workflow. the RSU computation and memory resources constrain task execution. Fig. 3 further specifies which actor performs each stage of the collaboration. First, the LLM agent on V 1 detects that its local observation is insufficient and sends a task request to the RSU. The RSU agent then decomposes the request, identifies the missing observation required for the turning decision, and discovers candidate agents on nearby vehicles. Next, the RSU coordinator and network controller determine the information source and message configuration according to the current design and resource conditions. The selected helper agent uses its local observation and LLM to generate task-relevant information and sends the resulting message to the RSU. Finally, the RSU agent incorporates the received information into the current task context. If the information remains insufficient, it initiates another interaction; otherwise, it sends the safety recommendation to V 1 , which executes or postpones the left-turn action. In simulations, an interaction is specified by the selected source agent and the amount of task-relevant information included in its message. The joint design additionally de- termines the wireless resources allocated to deliver the se- lected message. We use successful task completion as the task utility and report its average completion rate. A task is completed only when the selected information is sufficient for the safety decision and can be delivered and processed within the prescribed end-to-end deadline, which captures the latency constraint of the V2X task. The source and message decisions, as well as wireless resource allocation, are subject to the available uplink bandwidth and the RSU computation and memory constraints. We compare three designs under the same physical topology. The agent-only design adapts the source agent and message content with fixed wireless resource allocation. The wireless-only design fixes the source agent and message content while adapting the wireless resource allocation. The joint design adapts the source and message together with the wireless resource allocation while accounting 7 Fig. 4.Task completion rate under two bottleneck regimes. Under the information-limited regime (left), Îș measures the relative decision relevance of the key agent; under the bandwidth-limited regime (right), B max denotes the available uplink bandwidth. for the RSU execution constraints. Fig. 4 examines two communication regimes. In the information-limited setting, we define the key agent as the candidate agent whose observation is most relevant to the current safety decision. The key-agent importance ratio Îș characterizes the contribution of this agent relative to the other candidate agents. As Îș increases, selecting the key agent becomes more important for successful task completion. The wireless-only design cannot exploit this difference because its information source is fixed. The agent-only and joint designs can select the more useful source, while the joint design also accounts for the wireless feasibility of the resulting message. In the bandwidth-limited setting, the useful source is fixed and the available uplink bandwidth B max is varied. The bottleneck therefore shifts to message delivery. Increasing B max improves the support available to the selected interaction, while the joint design can further adapt the transmitted information amount to the available bandwidth. Fig. 5 examines the execution constraints at the RSU. The joint design achieves a higher task completion rate than the agent-only and wireless-only designs under both compute- and memory-limited settings. The RSU computation and memory demands are interaction-dependent: changing the source or message amount changes the information processed at the RSU and hence the resulting execution workload. The pro- cessing budget specifies the computation available before the task deadline, while the memory budget limits the amount of execution memory available at the RSU. The agent-only design may select an informative interaction whose resulting workload exceeds these constraints because execution feasi- bility does not enter its interaction decision. The wireless- only design can improve delivery of the predefined interac- tion but cannot change the computation or memory demand created after reception. The joint design can instead reduce the message amount or select another source when resources become limited, thereby adapting the generated workload to the available RSU resources. Fig. 5.Compute and memory bottlenecks under different agentânetwork designs (OOM is out of memory). V. CONCLUSION This article examined network-constrained LLM-based MASs from a joint agent and network design perspective. We showed that agent-side decisions and network operation should be jointly designed across interaction scheduling, infor- mation transfer, topology configuration, workload execution, and cross-layer trust. The V2X case study illustrated how jointly adapting interaction and resource decisions improves task completion under communication and edge-resource con- straints. These results motivate future wireless systems to coordinate agent collaboration with network connectivity and edge resources according to task requirements. REFERENCES [1] L. Wang et al., âA survey on large language model based autonomous agents,â Frontiers of Computer Science, vol. 18, no. 6, Art. no. 186345, 2024. [2] Q. Wu et al., âAutoGen: Enabling next-gen LLM applications via multi- agent conversation,â in Proc. Conf. Lang. Model. (COLM), 2024. [3] F. Jiang, Y. Peng, L. Dong, K. Wang, K. Yang, C. Pan, D. Niyato, and O. A. Dobre, âLarge language model enhanced multi-agent systems for 6G communications,â IEEE Wireless Commun., vol. 31, no. 6, p. 48â55, Dec. 2024. [4] M. Xu, D. Niyato, J. Kang, Z. Xiong, S. Mao, Z. Han, D. I. Kim, and K. B. Letaief, âWhen large language model agents meet 6G networks: Perception, grounding, and alignment,â IEEE Wireless Commun., vol. 31, no. 6, p. 63â71, Dec. 2024. [5] X. Cheng, J. Zhang, N. Ding, N. Li, Y. Li, T. Wu, W. Xu, J. Zhang, and Q. Sun, âIntegrated AI and communications: A two-way catalysis toward 6G and beyond,â J. Commun. Inf. Netw., vol. 10, no. 3, p. 191â200, Sep. 2025. [6] Y. Hu, X. Pang, X. Qin, Y. C. Eldar, S. Chen, P. Zhang, and W. Zhang, âPragmatic communication in multi-agent collaborative perception,â IEEE Trans. Pattern Anal. Mach. Intell., vol. 48, no. 8, p. 9279â9296, Aug. 2026, doi: 10.1109/TPAMI.2026.3680062. [7] K. Zhang, H. He, S. Song, J. Zhang, and K. B. Letaief, âCommunication- efficient distributed on-device LLM inference over wireless networks,â IEEE J. Sel. Topics Signal Process., vol. 19, no. 7, p. 1301â1317, Oct. 2025. [8] H. Seo, J. Seon, J. Y. Kim, M. Bennis, W. Choi, and D. I. Kim, âReasoning-native agentic communication for 6G,â arXiv:2602.17738, 2026. [9] G. Zhang et al., âCut the crap: An economical communication pipeline for LLM-based multi-agent systems,â in Proc. Int. Conf. Learn. Repre- sentations (ICLR), 2025. [10] G. Zhang, Y. Yue, X. Sun, G. Wan, M. Yu, J. Fang, K. Wang, T. Chen, and D. Cheng, âG-Designer: Architecting multi-agent communication topologies via graph neural networks,â in Proc. Int. Conf. Mach. Learn. (ICML), PMLR, vol. 267, p. 76678â76692, 2025. [11] W. Li, D. Qiao, B. Wang, X. Wang, W. Yin, H. Shen, B. Jin, and H. Zha, âMulti-agent credit assignment with pretrained language models,â in Proc. 28th Int. Conf. Artif. Intell. Statist. (AISTATS), PMLR, vol. 258, p. 1945â1953, 2025. 8 [12] D. Lee, M. Tiwari, and B. Miranda, âPrompt infection: LLM-to-LLM prompt injection within multi-agent systems,â in Computer Security. ESORICS 2025 International Workshops, LNCS, vol. 16232, p. 511â 520, Springer, Cham, 2026, doi: 10.1007/978-3-032-16092-8_28. [13] M. Costa, B. Köpf, A. Kolluri, A. Paverd, M. Russinovich, A. Salem, S. Tople, L. Wutschitz, and S. Zanella-BĂ©guelin, âSecuring AI agents with information-flow control,â arXiv:2505.23643, 2025. [14] Y. Li, D. Ma, Z. An, Z. Wang, Y. Zhong, S. Chen, and C. Feng, âV2X- Sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving,â IEEE Robot. Autom. Lett., vol. 7, no. 4, p. 10914â 10921, Oct. 2022. [15] J. Zhang, S. Hu, C. Lu, R. Lange, and J. Clune, âDarwin Gödel machine: Open-ended evolution of self-improving agents,â in Proc. Int. Conf. Learn. Representations (ICLR), 2026.