SCNOC-Agentic: A Network Operation and Control Agentic for Satellite Communication Systems

Abstract

Large language models (LLMs) have demonstrated powerful capability to solve practical problems through complex step-by-step reasoning. Specifically designed LLMs have begun to be integrated into terrestrial communication networks. However, relevant research in the field of satellite communications remains exceedingly rare. To address this gap, we introduce SCNOC-Agentic, a novel architecture especially designed to integrate the management and control of satellite communication systems in LLMs. SCNOC-Agentic incorporates four components tailored to the characteristics of satellite communications: intent refinement, multi-agent workflow, personalized long-term memory, and graph-based retrieval. Furthermore, we define four typical real-world scenarios that can be effectively addressed by integrating with LLMs: network task planning, carrier and cell optimization, fault analysis of satellites, and satellite management and control. Utilizing the SCNOC-Agentic framework, a series of open-source LLMs have achieved outstanding performance on the four tasks under various baselines, including zero-shot CoT, CoT-5, and self-consistency. For example, qwen2.5-70B with SCNOC-Agentic significantly improves the parameter generation accuracy in the network task planning task from 15.6% to 32.2%, while llama-3.3-70B increases from 16.2% to 29.0%. In addition, ablation studies were conducted to validate the importance of each proposed component within the SCNOC-Agentic framework.

1. Introduction

AI algorithms, exemplified by large language models (LLMs), are effectively applied across various vertical domains, particularly in communication network applications of space–air–ground–sea integrated networks (SAGSINs). In terrestrial communications, research on communication LLMs [1,2,3,4] has explored their integration with 5G networks [5,6], which could facilitate the development of LLMs for satellite-terrestrial networks. Benefiting from the coverage capability of a single communication satellite far exceeding that of ground base stations, satellite communications have emerged as an effective means to meet the growing demands of 6G and Internet of Everything (IoE) applications [7,8]. However, research on LLMs aimed at supporting ubiquitous satellite communication network services remains extremely limited.

In telecommunications, LLMs (e.g., GPT series [9,10], Qwen [11,12] series) have been extensively introduced to address various challenges within network systems. For instance, by designing traffic characterization and adjusting network output structures, these models facilitate a deeper understanding of network traffic [13,14,15]. Network operation and maintenance are also facilitated by the use of prompt engineering, network design, and copilot services [16,17,18]. Moreover, communication network security is enhanced by identifying network attacks and generating response plans [19]. Compared with traditional task-specific models, “LLM-powered networks” offer additional capabilities to communication systems. These include achieving intent understanding, altering the interaction methods between networks and operators, and enabling automated network operations. In addition, compared to terrestrial communications, satellite communications rely more heavily on the intentions of operations and maintenance personnel for network management, making applications based on LLMs in satellite communications more focused on network management and control aspects.

Despite the demonstrated capability of LLMs in executing complex step-by-step reasoning to solve problems, recent research in satellite communications has not kept pace with advances in model inference capabilities and agent functionalities. Specifically, there is no established definition or framework on how LLMs can be effectively integrated into satellite communication networks to address practical challenges within this domain, such as satellite networking planning [20,21,22], network management [23,24], traffic scheduling [25,26], and resource allocation [27,28,29,30]. Currently, there is also an absence of evaluation regarding whether LLMs can resolve management and control issues within satellite communication networks through step-by-step reasoning to derive solutions, alongside generating reliable parameters via agent invocation. Taking network construction in satellite communications as an example, when communication tasks for a specific region arise, it becomes necessary to quickly establish a satellite communication network tailored to these requirements, ensuring communication network coverage for that area. However, current practices in the operational management of satellite networks still rely on manually entering task parameters. This approach suffers from two main disadvantages: (i) Firstly, it requires high levels of expertise and reliability from personnel, which cannot guarantee the optimization of the parameters. (ii) Secondly, the networking process is time-consuming and requires several minutes to complete all procedures. Given the potential urgency of satellite communication tasks, such minute-level networking operations could result in missed communication opportunities.

To address these challenges, we propose a specially designed agentic framework for satellite communication operations and management, SCNOC-Agentic, as shown in Figure 1. This framework is designed for the application of LLMs in satellite communication systems, encompassing four critical components and making full use of the capabilities of planning, memory, and tool usage within the agentic method. In the SCNOC-Agentic framework, leveraging the intent parsing capabilities of LLMs, we have achieved a complete closed-loop process from communication requirements to the generation and distribution of network configuration parameters, realizing the goal of rapid and agile response. Looking ahead, we believe that the SCNOC-Agentic architecture can serve as the control center in satellite communication systems, capable of achieving full-process management and control of satellites, terminals, and ground stations.

The contributions of this work can be summarized as follows:

• We propose the SCNOC-Agentic framework for the operation and control of satellite communications, firstly marking the implementation of the LLMs agent architecture into this domain, which enhances the capability for rapid and agile response.
• In SCNOC-Agentic, we introduce four components specifically designed for the characteristics of the satellite communication field: intent refinement, multi-agent workflow, personalized long-term memory, and graph-based retrieval. The performance influence of each component is evaluated through ablation experiments.
• We design four typical scenarios for applying LLMs in satellite communications: network task planning, carrier and cell optimization, fault analysis of satellites, and satellite management and control. The integration of the SCNOC-Agentic framework within these scenarios is elaborated. Comparative experiments against current state-of-the-art general models and agents show that SCNOC-Agentic achieves superior performance across these four scenarios.

The remainder of this paper is organized as follows. Section 2 describes the related works. Section 3 details the proposed SCNOC-Agentic framework, which comprises four key components. Subsequently, Section 4 defines the scenarios for four typical applications. Section 5 describes the experimental results of SCNOC-Agentic in these four typical scenarios. Finally, Section 6 concludes the paper.

2. Related Work

2.1. AI-Agentic

LLMs serving as foundation models (FM), such as GPT, Qwen, Llama, and others, have demonstrated a powerful capability to perform complex step-by-step reasoning to solve practical problems [31,32,33,34]. This advancement is attributed to the in-depth research into AI agentic work. To enable LLMs to achieve more accurate and reliable responses when tackling complex real-world issues, it is generally necessary to integrate these models within specifically designed agentic systems that can access external tools such as database queries, proprietary model invocations, search engines, etc. An agentic system typically comprises three modules: planning [35,36], memory [37], and tool use [38,39], where the LLM plays a crucial role in decision-making and reasoning processes [40]. Furthermore, pioneering efforts to explore agents for general tasks, including AutoGPT and AutoGen [41], are designed to address a broad spectrum of general tasks by dissecting user inquiries into actionable components. These advancements signify a significant leap towards achieving versatile task automation facilitated by sophisticated AI agents.

In addition, researchers have developed a variety of building blocks and design patterns for different applications. This enables LLMs to solve complex problems by generating a series of intermediate steps, where each step addresses a simpler sub-problem, thereby progressively advancing towards the final solution [42]. For example, in [32], the Chain-of-Thought (CoT) prompting method was proposed. In [43], CoT-SC was introduced, which aggregates multiple parallel responses from CoT to produce more accurate responses. The self-refining method proposed in [44] allows for iterative self-reflection aimed at correcting errors made in previous attempts. Furthermore, LLM-Debate, introduced in [45], enables different LLMs to debate among themselves, leveraging diverse perspectives to arrive at superior solutions. In [46], Intelligent Go-Explore was presented, which facilitated the generation and integration of various responses to better explore potential solutions. Additionally, other crucial building blocks for agentic systems include prompting techniques [47,48], reasoning methodologies [34], reflection mechanisms [49], developing new skills for embodied agents [50], external memory and Retrieval-Augmented Generation (RAG) [51], assigning distinct roles to FM modules within the agentic system and enabling their collaboration [52,53,54], among others. These elements collectively contribute to enhancing the capability and versatility of AI agents to tackle complex tasks in various domains.

The design of LLM-based agent systems needs to be guided by industry-specific problems. Currently, such systems have shown significant potential in applications in domains such as finance [55], education [56], and medical diagnostics [54,57]. However, the field of satellite communications has not kept pace with the advancements in the reasoning capabilities of LLMs and the latest progress in agents’ abilities. Current research on applying LLMs to satellite communication tasks remains limited, with insufficient exploration of practical integration frameworks between satellite systems and agentic architectures.

2.2. LLM for Communication

Current research in the communication field has explored applications of LLMs. For instance, in [58,59], potential application directions of LLMs in communications are discussed. Specifically, in [60], Maatouk et al. introduced TeleQnA, a benchmark dataset to evaluate LLMs on telecom knowledge. In [61], Karim et al. proposed SPEC5G, a knowledge-based network protocol Q&A system. In [62], Miao et al. presented NetEval, which focuses on training LLMs for network operations. In [16], Wei et al. introduced an intention-based framework using LLM agents to translate network configurations and ensure network operation. In [63,64], approaches to training specialized LLMs for communications with small, curated text datasets were proposed. In [65], Wang et al. introduced NetAssistant, an LLM-based method for network fault diagnosis that selects appropriate troubleshooting processes based on user intent. In [66], Chen et al. focused on storing historical fault analysis records in a vector database, achieving cloud fault root cause analysis RCACopilot via LLM-based retrieval and analysis. In [67], Yuan et al. proposed UniST, a user mobility modeling method based on network-LLM. Similarly to the use of general LLMs, research on LLMs in the communication field focuses primarily on knowledge-based dialogue and Q&A to provide copilot services for daily maintenance.

Furthermore, research on multi-modal LLMs has also been conducted in the field of communications. For example, the performance of 6G [68], networks [17], and specific issues [14] in the communication domain has been enhanced by incorporating visual information. Simultaneously, multi-modality in the field of communications is not limited to images and text; network traffic also constitutes a significant modality. For example, in [19], Wang et al. proposed ShieldGPT, a feature engineering-based network LLM that utilizes text representations of binary traffic data, which is capable of interpreting slowbody attack behaviors and generating diverse defense instructions. In [69], Wang et al. adopted a unidirectional Mamba architecture to propose an efficient pre-trained traffic classification model, NetMamba. In [17], Wu et al. introduced NetLLM, which designs a multimodal encoder, removes the default language model head, and adds a trainable network head tailored for network tasks. In [8], Sun et al. proposed satellite orchestration in 6G via AI-native frameworks and multi-agent DRL to enhance edge computing efficiency and explained the role of AI agents in achieving cross-domain collaboration. In [70], Rong et al. explored LLMs for intelligent control in 6G TN-NTN to address resource management, interference cancellation, and handover challenges. In [71], Hao et al. proposed ToolkenGPT, which represents each network tool as a new “toolken” and allows the LLM to learn the embedding representations of these toolkens, thus directly generating toolkens to trigger tool invocations.

3. Overview of SCNOC-Agentic

This section introduces the proposed SCNOC-Agentic framework designed for satellite communication operation and control, as shown in Figure 1, elaborating on its four key components.

3.1. Intent Refinement

LLMs inherently possess the ability to understand and parse input intent [16], enabling a more effective interaction with human intentions. However, the intent parsed by LLMs is not always precise. To address this, we introduce an intent refinement (IR) component in SCNOC-Agentic, inspired by the Intent-Driven Networks (IDN) approach [72]. The IR component validates and refines the user input requirements. As shown in Figure 2, using a satellite beam adjustment task as an example, when the input is “Adjust the pointing beam of Ku1 of the TT03 satellite to the Hawaii region,” the IR component corrects the parsed intent parameters to their accurate values. This significantly improves the accuracy of intent translation and simplifies traditional intent management and network operation processes.

The IR component encompasses several steps, including intent translation, template matching, and conflict detection. Among these, the intent template matching step facilitates the alignment of parameter templates. It ensures the correctness of both the number of parameters and their keywords by aligning the parameter names generated by the model with the actual parameters issued through intent template matching. The intent translation step converts the intent into parameters that can be understood by satellite communication agents. For instance, “Query the beam coverage of the Hawaii region” would translate “Hawaii region” into specific latitude and longitude coordinates [20 N, −157 W]. The conflict detection step verifies resource conflicts, such as checking whether a specific satellite beam frequency resource is already occupied when applying for it, thereby preventing conflicts.

Notably, we have integrated Named Entity Recognition (NER) [73] into the IR component to handle cases where users employ abbreviated terms instead of full names, such as using “TT03” for ”TianTong-2 Satellite 03”. By incorporating Named Entity Recognition for matching and correction, the actual system’s robustness and usability are significantly enhanced.

3.2. Multi-Agent Workflow

Multi-agent collaboration [52,53,54] is a crucial design for LLMs to address complex satellite communication tasks. In real-world satellite communication scenarios, LLMs generally need to invoke different agents through multiple steps to resolve various issues separately. For example, when fulfilling the requirements of network task planning, resource querying and allocation are typically performed first, followed by network construction. Therefore, we have introduced a multi-agent workflow (MaW) component, as shown in Figure 3, to establish a fixed workflow for handling complex network management and control tasks.

The MaW component is designed to handle functional requests (as opposed to question–answering) in satellite communications. It operates by sequentially executing a series of nodes, making it suitable for automated satellite operations, network formation, and resource optimization. At its core, the MaW component consists of nodes, each of which is an independent agent with a specific function. These agents perform critical operations such as task analysis of satellite communications, resource management, and network control. By connecting the outputs of different agents to subsequent inputs, multiple agents can be linked together, enabling plug-in integration and workflow orchestration.

The multi-agents are called through predefined connection relationships. For example, a beam adjustment agent is restricted from invoking a traffic scheduling model but is allowed to invoke a resource query agent. When predefined tasks arise, they are decomposed into workflows, thereby enhancing the handling of intricate satellite communication operations and control challenges. Furthermore, the MaW incorporates the invocation of specific small-scale models. All agents can leverage specialized models within satellite communications, such as traffic prediction models, resource allocation models, and anti-jamming models, to facilitate decision-making support. The distinction between the MaW and CoT lies in the fact that the CoT method merely provides templates for LLM agent invocations, whereas the MaW imposes stringent constraints on system-level agent invocations. The MaW component significantly improves the accuracy of multistep agent invocations for complex tasks within the SCNOC-Agentic framework.

3.3. Personalized Long-Term Memory

The uniqueness of satellite communications lies in its diverse user base, where different users, such as maritime, vehicular, and airborne users, exhibit distinct behavioral patterns. To address this, we propose a personalized long-term memory (PLM) component that personalizes the recording of usage habits and behaviors for each type of user. During model inference, this component provides context-specific prompts to the model, leveraging advances in long-context LLMs. The PLM component consists of two parts: preference and memory. Preference focuses on the rules in the field of satellite communications, and memory focuses on historical user data.

As shown in Figure 4, the PLM component sets different preferences for different users and stores the memory of previous usages for each user. This design can have notably beneficial effects in the field of satellite communications. For example, maritime users typically rely on a fixed number of satellites and prefer to use the Ku/Ka band. Given that various types of satellite users have relatively stable behaviors and preferences, incorporating and customizing the PLM component allows LLMs to distinguish between different types of users. Consequently, this customization facilitates the provision of valuable insights to assist LLMs in analysis and decision-making.

During the reasoning process, the storage and retrieval process of the PLM component operate as follows: (1) Storage: The system automatically identifies and records personalized information provided by users in conversations, such as mobile users or ship-based users. Long-term memory in satellite communications should be mutually isolated among users. (2) Retrieval: Long-term memory can be retrieved either through user prompts or at designated steps within the MaW component. When users request retrieval of relevant long-term memory, SCNOC-Agentic generates the final response by summarizing the personalized information as prior knowledge.

3.4. Graph-Based Retrieval

In light of the complexity of satellite communication systems, to enhance the accuracy of invocation more reliably, we have constructed the graph-based retrieval (GBR) component utilizing the Graph-RAG approach [73]. The abundance of research on Graph-RAG underscores its aptitude for encoding heterogeneous and relational information intrinsic to satellite communication systems, making it highly suitable for practical applications. By integrating real-time data retrieval capabilities, the Graph-RAG system can effectively enhance the functionalities of SCNOC-Agentic.

Furthermore, recognizing the privacy concerns associated with certain satellite communication systems, where satellites, beams, and terminal names are often absent from the pre-training datasets of LLMs, the introduction of the GBR component enables SCNOC-Agentic to identify named entities not encountered during the LLMs’ pre-training phase. Specifically, our GBR component is built upon an accumulated corpus of satellite communication documents, including satellite system design documents, user manuals, educational materials on satellite communications, and other relevant texts. During the graph construction phase, LLMs extract entities and relationships from structured texts, visually representing these connections in a graph format. This allows for joint modeling of entities and their relationships during the retrieval process. In the inference phase, the vectors generated based on user queries facilitate the retrieval of related entities and corresponding relationships, indexing knowledge blocks pertinent to queried entities, thereby enhancing the performance of the LLMs.

4. Scene and Problem Formulation

In this section, we provide detailed definitions of agents for four typical scenarios in which SCNOC-Agentic is applied and describe how LLMs can be integrated with these satellite communication scenarios.

4.1. Network Task Planning

With the development of satellite internet, the scale of space-based networks is further expanding, necessitating a higher level of abstraction for traditional satellite resource allocation and network construction methods. In response to this, our project proposes the integration of LLMs based on the SCNOC-Agentic architecture with intention-driven networks (IDN) [72], as illustrated in Figure 5. This integration aims to facilitate intent-based task planning in satellite networks.

The network task planning task for satellite communications based on the SCNOC-Agentic framework is defined as follows: understand the communication intent requirements, then complete the allocation of satellite, earth station, and terminal resources, and finally complete the network construction. For example, if the user inputs 2025.2.10 to ensure a 30M TDMA communication network in Hawaii, SCNOC-Agentic can understand and parse the intent input and perform accurate agent selection and parameter generation step by step. Specifically, under the given conditions of a satellite communication system, SCNOC-Agentic utilizes the proposed PLM component to prompt on the basis of the user’s previous task requirements. Currently, it retrieves relevant knowledge bases using the proposed GBR component to generate prompts. Intent parsing and validation are performed using general LLMs, as outlined in Section 3.1. The proposed MaW component formulates the requirements for satellite communication tasks into a multistep execution workflow. In each step, different agents are invoked by the LLM to execute corresponding functions.

As shown in Figure 5, the workflow of the network task planning task is divided into three steps, with each step invoking the corresponding agent for processing: (1) Network resource query: This initial step involves querying the available resources at the specified networking location (e.g., Hawaii region) for network task planning. SCNOC-Agentic generates the necessary parameters and invokes the appropriate agent. The interface definitions utilize standard MCP and Function Call protocols. The input format is defined as Action input:[“input.networktype ”: 2, “input.latitude”: 40, “input.longitude”: 116, “input.time”: datetime.now()]. The query results include available satellites, the area of beams cover, ground stations, and terminals within the region, formatted as {[satelliteId1, satelliteId2], [beamId1, beamId2], [terminalId1, terminalId2, terminalId3]}. (2) Network resource allocation: Following the resource query, this step involves selecting resources based on the query results. It consists of two parts: selecting satellites, beams, and transponders and choosing available ground stations and terminals. Subsequently, frequency resources from the selected satellites and beams are allocated (e.g., dividing 30M from a 100M transponder resource while ensuring no time conflicts), similar to satellite communication resource allocation studies [27,28]. (3) Network construction: In the final step, SCNOC-Agentic generates the parameters for network construction and sends them to the network management system to complete the construction of the satellite communication network involving satellites, beams, and user terminals. This structured approach ensures efficient and effective management and control of satellite communication networks, tailored to meet specific user requirements.

Compared to traditional satellite communication network task planning, SCNOC-Agentic-based network task planning offers additional gains by facilitating an end-to-end process from user intent input to network construction. Traditional methods [74] exclusively encompass resource planning in terms of frequency and time slots (Step 2 in Figure 5). In contrast, SCNOC-Agentic network task planning includes satellite and beam selection, along with the generation of networking parameters.

4.2. Carrier and Cell Optimization

The next generation of broadband communication satellites and satellite internet node satellites generally adopt onboard processing systems, equipped with phased array antennas. Optimization of the carrier and cell of phased array antennas is a critical technology for achieving low-latency satellite communications for the ubiquitous Internet of Everything (IoE). Phased array antennas significantly enhance the spatial diversity of the system, enabling time division coverage of multiple dispersed areas on the Earth’s surface within certain periods, thereby covering more regions with fewer beam-hopping events. Furthermore, methods such as [29,30] provide communication capabilities that match service demands by adjusting the dwell time of beams in different areas within a period.

In the carrier and cell optimization task, SCNOC-Agentic compensates for link disparities to ensure communication quality by adjusting the Modulation and Coding Scheme (MCS) instead of slot allocation, as shown in Figure 6. This approach fundamentally addresses limited-response problems. By predefining M sets of carrier groups, each representing an MCS, SCNOC-Agentic selects among these to reconcile transmission capability differences caused by varying antenna aperture sizes and output powers of ground-user stations. Different carriers are configured with distinct MCSs to accommodate various types of stations. By adjusting the MCSs, SCNOC-Agentic aims to guarantee access for most station types while maximizing the bandwidth utilization. Moreover, it optimizes link impacts from phased array antennas operating in high-frequency bands such as Ku/Ka and Q/V (e.g., mitigating rain fade). SCNOC-Agentic achieves carrier and cell optimization by generating parameters for MCS adjustments and terminal priority settings, but it cannot calculate the optimal End-to-End Latency (E2EL). Hence, three evaluation metrics are employed: Agent Select Accuracy (AGA), Parameter Generation Accuracy (PGA), and Average End-to-End Latency (E2EL), with E2EL reflecting the average delay of all terminals under specific carrier templates and priority settings. Given that carrier and cell optimization is NP-hard, making the calculation of optimal E2EL unfeasible, comparisons are made using a fixed number N of terminals to evaluate the relative performance.

4.3. Fault Analysis of Satellite

The fault analysis of a satellite via SCNOC-Agentic essentially constitutes a multi-classification problem, thereby enabling the formulation of tasks as multiple-choice questions. Utilizing SCNOC-Agentic not only facilitates the identification of single-point equipment failures, such as satellite platform and payload malfunctions but also extends to the analysis of network-level and link failures, including inter-satellite link and feeder link failures. This capability is attributed to the SCNOC-Agentic’s foundation on LLMs, which enables simultaneous analysis of onboard telemetry data, log data, and parameter data from various devices—capabilities beyond the reach of traditional single-fault analysis models.

We employed a self-developed LEO satellite simulation center to construct experimental scenarios by receiving telemetry data from payload unit boards. This ground-based simulation system is designed as a payload controller for LEO satellite constellations, featuring integrated processing simulations for two satellites. It emulates the operational status of a satellite communication network, comprising two satellites in 500 km nearly polar orbits, and supports both payload remote control and telemetry functions.

Artificially setting the payload fault parameters, we simulated four types of payload anomalies: control board failure (main control unit payload), satellite–ground exchange unit failure, modem board unit failure, and inter-satellite exchange unit failure. A normal payload state was also configured for comparison. For each state, 90 min of telemetry data were collected to establish the experimental dataset, as shown in

Table 1. Telemetry information of satellite payload fault status.

Telemetry Status	Fault Payload	Data Size
Normal	-	12,124
Fault state 1	main control unit	5312
Fault state 2	satellite-ground switching unit	5769
Fault state 3	modem board card unit	4793
Fault state 4	inter-satellite switching unit	4902

. For the fault analysis of the satellite task, the evaluation primarily adopted the top-k recall rate (Hit@k, where k = 1, 3) as the key metric, with the Hit@k measure accuracy by verifying whether the correct fault analysis appeared within the top k predictions.

4.4. Satellite Management and Control

The implementation of satellite management and control tasks based on the SCNOC-Agentic framework involves controlling the satellite platform and its payloads. As shown in Figure 7, first, SCNOC-Agentic interprets the intentions of the control commands from maintenance personnel and performs the correct agent selection. Subsequently, the selected agent translates these intentions into a set of instructions and uploads them to the satellite. Finally, adjustments to the satellite are verified through telemetry data. Typical control command intentions include: “Adjust Ka beam pointing to Hawaii”, “Activate inter-satellite laser payload”, “Switch X satellite phased array mode to stare”, and “Switch X satellite to secure communication mode”. Maintenance personnel customize common command intents based on the SCNOC-Agentic architecture. The PLM module provides prompts tailored to different individuals, enabling various methods of satellite operation and control, significantly enhancing the efficiency of management. This task is evaluated through ASA and PGA.

In satellite and network management, smaller scales allow focus on detailed parameter configurations of individual satellites. However, as the satellite internet expands, along with the broadening operational dimensions, managing all parameter details becomes increasingly impractical. Consequently, SCNOC-Agentic represents a higher level abstraction of traditional single-satellite parameter-based management and control, emphasizing intent understanding for management. Furthermore, the proposed SCNOC-Agentic framework is designed to adapt to more complex network configurations in the future, with the continuous development of LLMs enhancing intent-driven capabilities.

5. Experiments

In this section, the proposed SCNOC-Agentic framework is validated on various general-purpose LLMs, including the latest versions of Llama3.3-70B, qwen2.5-70B, GPT-4o, and GPT-3.5-turbo-0125. All models were assessed using zero-shot CoT [32], by default, with the temperature parameter set to 0. Various baseline approaches were used, including zero-shot CoT, which uses the prompt “let us think step by step,” and the CoT prompting technique with five examples (CoT5) [33]. Regarding self-consistency (SC) [43], the accuracy was determined based on the majority decision of 20 generations. All baselines completed experiments on four typical satellite communication operation and control application scenarios presented in Section 4, along with ablation studies on components. Moreover, on the basis of the experimental results, we discuss how LLMs can be practically applied to satellite communication operation and control systems.

5.1. Main Results

We evaluated the performance of the SCNOC-Agentic framework under various baselines.

Table 2. Comparative experiments on network task planning and carrier and cell optimization tasks.

Model	Network Task Planning			Carrier & Cell Optimization
	ASA	PGA	AoP	ASA	PGA	E2EL (ms)
General LLMs (default with zero-shot CoT)
GPT-4o	0.6736	0.2046	0.835	0.7511	0.5146	302
GPT-4o+CoT5	0.7321	0.2339	0.863	0.8164	0.5943	286
GPT-4o+SC	0.8101	0.3898	0.903	0.8449	0.7177	290
GPT3.5	0.6171	0.1754	0.835	0.7968	0.4577	343
GPT3.5+CoT5	0.6853	0.1968	0.846	0.7970	0.4962	318
GPT3.5+SC	0.7594	0.3294	0.887	0.8132	0.6037	294
Llama-3.3-70B-Instruct (default with zero-shot CoT)
RawAgent	0.6035	0.1559	0.812	0.6697	0.4608	346
RawAgent+CoT5	0.7029	0.2202	0.840	0.7784	0.5120	280
RawAgent+SC	0.7769	0.3664	0.863	0.7943	0.5753	285
SCNOC-Agentic	0.7029	0.3216	0.887	0.8101	0.6575	312
SCNOC+CoT5	0.7867	0.3762	0.927	0.8354	0.7335	260
SCNOC+SC	0.8491	0.4619	0.945	0.8670	0.7683	277
Qwen-2.5-70B-Instruct (default with zero-shot CoT)
RawAgent	0.6346	0.1617	0.829	0.6840	0.4703	314
RawAgent+CoT5	0.7243	0.2358	0.843	0.7689	0.5151	280
RawAgent+SC	0.7638	0.3586	0.872	0.8006	0.6006	274
SCNOC-Agentic	0.7185	0.2904	0.854	0.8575	0.6797	298
SCNOC+CoT5	0.7828	0.3411	0.921	0.8829	0.7145	273
SCNOC+SC	0.8276	0.4249	0.939	0.9082	0.7841	264

presents the comparative results of the proposed SCNOC-Agentic framework with different general LLMs and prompting methods on network task planning and carrier and cell optimization. The best two results are highlighted, with the best result bolded, underlined, and marked in red, while the second-best is underlined and marked in blue.

As observed in Table 2, the SCNOC-Agentic-based framework achieved optimal and suboptimal results across various tasks. In the network task planning task, Llama-3.3-70B+SCNOC-Agentic+SC demonstrated the best performance in all metrics. The utilization of the SCNOC-Agentic architecture resulted in improvements of 7.2% and 9.5% in the AGA (0.8491 vs. 0.7769) and PGA (0.4619 vs. 0.3664) metrics, respectively. Compared to the baseline using CoT by default, the SCNOC-Agentic architecture still yielded enhancements of 9.9% and 16.7% in the AGA (0.7029 vs. 0.6035) and PGA (0.3216 vs. 0.1559) metrics, respectively. Similarly, in the carrier and cell optimization task, the SCNOC-Agentic architecture outperformed the Llama and Qwen models under RawAgent, delivering superior AGA and PGA performance, as well as reduced channel latency. These results collectively substantiate the efficacy of the SCNOC-Agentic framework.

Meanwhile, although we observe that the proposed SCNOC-Agentic has achieved significant improvements over state-of-the-art general LLMs and the GPT-4o model, we believe that the current agents framework still falls short of practical application in satellite communication systems, particularly in network task planning and carrier and cell optimization tasks. For practical use, it is essential to ensure that the performance of AGA and PGA is as high as possible, with parameter generation being as reliable as possible. Therefore, more robust foundational models and engineering tricks or, alternatively, the design of a superior agent architecture are required.

Similarly, as shown in

Table 3. Comparative experiments on fault analysis of satellite and satellite management and control tasks.

Model	Fault Analysis of Satellite			Satellite Management and Control
	Hit@1	Hit@3	Hit@5	ASA	PGA
General LLMs (default with zero-shot CoT)
GPT-4o	0.190	0.315	0.725	0.8908	0.5285
GPT-4o+CoT5	0.214	0.360	0.796	0.9430	0.5761
GPT-4o+SC	0.246	0.404	0.818	0.9791	0.6238
GPT3.5	0.164	0.287	0.683	0.7952	0.4809
GPT3.5+CoT5	0.182	0.373	0.742	0.8821	0.5285
GPT3.5+SC	0.227	0.388	0.794	0.9256	0.5761
Llama-3.3-70B-Instruct (default with zero-shot CoT)
RawAgent	0.171	0.269	0.693	0.8121	0.4285
RawAgent+CoT5	0.208	0.349	0.760	0.9169	0.5761
RawAgent+SC	0.203	0.370	0.781	0.9343	0.6714
SCNOC-Agentic	0.185	0.341	0.755	0.9430	0.7238
SCNOC+CoT5	0.224	0.404	0.796	0.9517	0.8714
SCNOC+SC	0.253	0.417	0.827	0.9691	0.9142
Qwen-2.5-70B-Instruct (default with zero-shot CoT)
RawAgent	0.177	0.272	0.684	0.8213	0.4285
RawAgent+CoT5	0.190	0.343	0.770	0.9082	0.5238
RawAgent+SC	0.216	0.362	0.783	0.9430	0.6714
SCNOC-Agentic	0.198	0.354	0.742	0.9704	0.7714
SCNOC+CoT5	0.238	0.394	0.805	0.9778	0.8190
SCNOC+SC	0.259	0.410	0.852	0.9865	0.8667

, we conducted comparisons using the same baselines and LLMs on the tasks of fault analysis of satellites and satellite management and control to fully evaluate the proposed SCNOC-Agentic framework. It can be observed that the proposed framework achieved the best and second-best results in both tasks. Particularly in the satellite management and control task, Qwen2.5-70B+SCNOC-Agentic+SC demonstrated exceptional performance with an ASA metric of 0.9865 and a PGA metric of 0.8667. This indicates that LLMs based on SCNOC-Agentic are well-suited for relatively simpler satellite management and control tasks. It is evident that SCNOC-Agentic outperforms the state-of-the-art GPT-4o and RawAgent in the same zero-shot CoT, CoT5, and SC settings, with significant improvements in both ASA and PGA metrics. This enhancement is attributed to the intent resolution capabilities provided by various components within SCNOC-Agentic. The MaW component improves the agent’s ability to accurately invoke functions, while the PLM and IR components significantly enhance parameter generation capabilities. The ablation experiments for each component will be discussed in Section 5.2.

In addition, we observed that the performance of Hit@1 in the fault analysis of the satellite task was merely 25.9%, which is far from satisfactory for practical applications. However, if the Hit@5 accuracy metric is adopted, the optimal result achieved by SCNOC-Agentic reached 85.2%. Therefore, for the deployment of the fault analysis of satellite task in real-world satellite communication systems, it is strongly recommended to utilize the best multiple recommendations provided by LLMs, with final confirmation by the operation and maintenance personnel. Finally, from a qualitative perspective, applying the agent capabilities of LLMs to the fault analysis of satellite tasks proved to be less effective compared to satellite management and control tasks. This discrepancy arises because fault analysis involves model-based reasoning and decision-making, whereas satellite management and control primarily concern the generation of interface parameters. Our findings, as outlined in Table 2 and Table 3, indicate that although current SCNOC-Agentic systems can effectively replace satellite management and control systems in practical applications, substantial advancements in LLM agent frameworks are still required for tasks involving computational reasoning in satellite communications before these systems can be confidently implemented in operational environments.

In this section, we conducted experiments to evaluate the importance of each component of SCNOC-Agentic under various scenarios, with the results presented in

Table 4. Ablation study on the importance of components on network task planning and carrier and cell optimization tasks.

Model	Network Task Planning			Carrier & Cell Optimization
	ASA	PGA	Ro	ASA	PGA	E2EL (ms)
Llama-3.3-70B-Instruct (default with zero-shot CoT)
w/o IR	-	0.2317	0.834	-	0.5082(↓)	330(↓)
w/o WAG	0.6629	0.2472	0.841	0.7991	0.5746	328
w/o PLM	0.6450	0.2191	0.825	0.7825(↓)	0.5284	322
w/o GBR	0.6912	0.2904	0.879	0.8029	0.6386	293
SCNOC-Agentic	0.7029	0.3216	0.887	0.8101	0.6575	321
Qwen-2.5-70B-Instruct (default with zero-shot CoT)
w/o IR	-	0.2241	0.849	-	0.5664	293
w/o WAG	0.6891	0.2598	0.831	0.7943	0.5341(↓)	346(↓)
w/o PLM	0.6639	0.2205	0.847	0.7879(↓)	0.5468	304
w/o GBR	0.7084	0.2745	0.862	0.8354	0.6802(↑)	284
SCNOC-Agentic	0.7185	0.2904	0.854	0.8575	0.6797	298

and

Table 5. Ablation study on the importance of components on fault analysis in satellite and satellite management and control tasks.

Model	Fault Analysis of Satellite			Satellite Management and Control
	Hit@1	Hit@3	Hit@5	ASA	PGA
Llama-3.3-70B-Instruct (default with zero-shot CoT)
w/o IR	0.179	0.303	0.693	-	0.5809(↓19.7%)
w/o WAG	0.184	0.315	0.719	0.8721(↓7.5%)	0.6455(↓10.8%)
w/o PLM	0.177	0.297	0.707	0.8847(↓6.4%)	0.6109(↓15.6%)
w/o GBR	0.190	0.338	0.736	0.9343(↓0.9%)	0.6761(↓6.6%)
SCNOC-Agentic	0.185	0.341	0.755	0.9430	0.7238
Qwen-2.5-70B-Instruct (default with zero-shot CoT)
w/o IR	0.180	0.301	0.710	-	0.6092(↓21.0%)
w/o WAG	0.187	0.318	0.706	0.9095(↓6.2%)	0.6761(↓12.4%)
w/o PLM	0.183	0.306	0.692	0.8908(↓8.2%)	0.6285(↓18.5%)
w/o GBR	0.195	0.320	0.714	0.9617(↓0.8%)	0.7238(↓6.2%)
SCNOC-Agentic	0.198	0.354	0.742	0.9704	0.7714

. The experiments were conducted under the default setting of zero-shot CoT, where ’w/o’ is an abbreviation for ’without’, such as ’w/o PLM’ indicating the exclusion of the PLM component within the SCNOC-Agentic framework while still utilizing the other three components. Arrows represent the performance gain(↑) or drop(↓) from component removal compared to baseline, as well as the percentage change compared to the baseline. The results of the ablation study demonstrate that all components contribute positively, highlighting the utility of each component in the design of SCNOC-Agentic. We observed that the removal of the IR component leads to a decline in PGA performance but has almost no impact on ASA. This is in line with our expectations. The reason is that the IR component enhances the accuracy of parameter selection by performing intent correction using named entity recognition. The role of the IR component is merely to correct parameters, thus it does not affect the choices made by the agent.

5.2. Components Ablation Study

Additionally, as shown in Figure 8, we calculated the performance drop in each component across four tasks. The removal of WAG or PLM components was observed to lead to a significant performance drop, indicating that incorporating WAG and PLM components is crucial for practical tasks. However, it was noted that eliminating the GBR component results in only a minor performance drop, suggesting that the introduction of satellite communication-related knowledge does not significantly enhance the performance of the LLM agents. However, the inclusion of the GBR component remains meaningful; for instance, when operators ask question-and-answer type questions (e.g., how many satellites are there in the TianTong constellation?), the GBR component can assist the LLM in generating more definitive responses by retrieving information from the knowledge base.

5.3. Discussion and Analysis

Compared with traditional algorithms: Compared to traditional feature learning models, the introduction of LLMs in the field of satellite communications brings not only performance improvements but also additional application scenarios, achieving objectives that small models cannot achieve. Meanwhile, the integration of LLMs enables comprehensive state management across the entire satellite network system, whereas smaller models are typically optimized for specific domains or tasks without considering the overall system comprehensively. Unlike small task-specific models, LLMs possess the capability to orchestrate tasks, allowing an end-to-end process from initial communication requirements to decision-making outcomes, thus significantly enhancing the efficiency of satellite communication operations and control. Furthermore, from the perspective of generalization, versatile LLMs can better adapt and transfer across various tasks and scenarios in satellite communications, a feat unachievable by traditional models.

Additionally, we propose that the architecture “LLMs + specific small models” represents the optimal approach for integrating LLMs with existing satellite communication systems and algorithms. This combination brings new benefits, rather than replacing traditional small models outright. Given that LLMs have clear limitations and may not be as precise or efficient in specific areas or tasks (such as traffic scheduling and resource allocation), they can invoke smaller models to handle these detailed tasks. For instance, during communication network formation, LLMs can first call upon smaller models for resource allocation, then proceed to issue network formation tasks once idle resources are identified. This synergy maximizes the strengths of both LLMs and small specialized models.

Benefits of SCNOC-Agentic: We posit that incorporation of SCNOC-Agentic into satellite communications operations and maintenance control yields significant enhancements, specifically: (1) SCNOC-Agentic advances typical communication scenario matching and intelligent parsing of communication intents based on intent-based methodologies. This elevates traditional manual conversion of communication requirements or the human designation of network parameters to an intelligent automated task analysis driven by LLMs and intent comprehension, transitioning from a “human-in-the-loop” to a “human-on-the-side, fully automated” operational paradigm. (2) After integration, SCNOC-Agentic streamlines information access for communication maintenance personnel, with LLMs acting as the “next-generation operating system” central to inputting requirements. Maintenance personnel interact solely with the Copilots system, issuing communication intents via LLMs to orchestrate comprehensive satellite communication system responses, such as querying available bandwidth over Hawaii, adjusting beam directions, or tallying beams covering the region. (3) Functioning as the control nucleus, LLMs adapt to the evolving state of satellite communication networks, coordinating subsystems and specialized “mini-models” for end-to-end management. This realizes a dynamic resource adjustment approach driven by service demand and optimized for user experience, effectively balancing resource allocation across the entire domain and optimizing resource utilization to enhance both satellite resource efficiency and service assurance capabilities. With the continuous advancement of LLMs, it holds the potential to become a critical technological pathway towards achieving autonomy in satellite networks.

In addition, each of the four application scenarios mentioned above is complex and deserves separate study. For example, constructing a satellite network involves various constellations, satellite types, antennas, and transponders. This paper only presents the simplest possible examples of how the LLMs’ agentic architecture might be applied in the field of satellite communications. Subsequent research needs to be developed based on different satellite communication systems, such as GEO and LEO, spot beam and multi-beam, etc. The details of satellite networks vary across different systems, making this a complex issue in itself. It is hoped that future research will explore these applications in greater depth.

6. Conclusions

In this paper, we introduced the SCNOC-Agentic framework, a large language model architecture specifically designed for the management and control of satellite communication systems. Within SCNOC-Agentic, we designed four components tailored to the characteristics of satellite communications: intent refinement, multi-agent workflow, personalized long-term memory, and graph-based retrieval. Subsequently, we conducted ablation experiments to evaluate the performance impact of each individual component. Furthermore, we defined four typical scenarios for the application of LLMs within satellite communications: network task planning, carrier and cell optimization, fault analysis of satellites, and satellite management and control. We provided detailed explanations of how SCNOC-Agentic could be applied in these scenarios, demonstrating superior performance compared to current state-of-the-art general models and agents.

It should be noted that the definable agent scenarios within the satellite communication domain extended beyond the four outlined in this study. Each of the four application scenarios mentioned above is complex and deserves a separate study. For example, constructing a satellite network involves various constellations, satellite types, antennas, and transponders. This paper presents only the simplest possible examples of how the LLMs agentic architecture might be applied in the field of satellite communications. Potential research areas within satellite communication management and control include user service assurance, fault analysis, traffic analysis, and situation generation, among others. For example, future work may focus on how to enable LLMs to understand communication protocols and assist in network management. In addition, the proposed SCNOC-agentic system is still limited by the workflow and prompting methods. Therefore, autonomous planning of multi-agents is a worthy research direction in the future.

In the future, subsequent research should continue to be conducted according to different satellite communication architectures, such as GEO and LEO, spot beam and multi-beam, etc. The details of satellite networks vary across different systems, making this a complex issue in itself. It is hoped that future research will explore these applications in greater depth.