This is typically divided into short-term and long-term components, often backed by vector databases (Xie et al., 2023) to support document and embedding retrieval. In general-purpose agents, memory may also store reasoning traces, self-critic logs, and state histories to support planning and reflection. In RAN-specific agents, this component must expand to handle structured time-series data, performance counters, and domain-specific logs. However, no standard exists for representing or exchanging these heterogeneous data types across agent implementations. Further, depending on how the agent is distributed—whether across the protocol stack or the cloud-to-edge continuum—memory allocation and access latencies may vary significantly. This necessitates standardized memory management schemas and access protocols that account for RAN's real-time constraints. Moreover, memory-augmented decision-making in RAN settings requires consistent traceability and auditability, which are yet to be addressed in existing implementations.

Agents use a variety of planning methodologies, ranging from chain-of-thought prompting (CoT) (Wei et al., 2022) to more sophisticated approaches of ReAct (Yao et al., 2023b), tree-of-thought (ToT) (Yao et al., 2023a), and graph-of-thought (GoT) (Besta et al., 2024). These strategies enable multi-step reasoning, task decomposition, and iterative refinement. However, while these approaches have shown promise in general domains, standard evaluation metrics for planning performance in RAN-specific tasks—such as latency impact, task completion reliability, and error recovery—are lacking. Moreover, agent planning capabilities often include self-reflection and multi-plan selection. Without standardized representations for plan structures, execution traces, and retry logic, it becomes difficult to compare or benchmark different agents under telecom-AI workloads.

Planning components often culminate in the invocation of external tools—such as network configuration APIs, monitoring scripts, or diagnostic utilities—to complete a task. Modern LLM agents can even generate tools on the fly, using in-context code synthesis. LLM-based code refactoring (Du et al., 2023; Thakur et al., 2023), network configuration generation (Xu et al., 2024; Dzeparoska et al., 2023), and policy translation (Mondal et al., 2023; Wang et al., 2023) are increasingly integrated into telecom use cases. However, without standardized tool registries, interface definitions, or capability descriptors, dynamically generated tools raise significant concerns around security, traceability, and compatibility. For example, the integration of techniques such as Shapley value calculation (Winter, 2002) can help assess the contribution of individual RAN elements (e.g., base stations, cell configurations, user load) toward network KPIs. Yet, the adoption of such tools requires standardized input formats, attribution models, and execution environments. Currently, the use of AI-generated tools in RAN management remains ad hoc, and their deployment poses integration risks without consistent operational semantics.

Function-level specifications are commonly expressed through standards such as OpenAPI 3.0 (OpenAPI Initiative, 2017), which provide a foundational semantic layer compatible with baseline LLMs. However, these specifications alone are insufficient for building fully functional and interoperable agent ecosystems in the RAN context. To realize a practical agent framework, it is essential to complement such semantic descriptors with concrete implementation guidelines, runtime requirements, and domain-specific execution semantics.

Development platforms such as NVIDIA's Aerial RAN CoLab Over-the-Air (ARC OTA) (NVIDIA, 2024a) and signal processing libraries such as PyAerial (NVIDIA, 2024b) represent critical infrastructure for accelerating the prototyping of agent capabilities. These platforms offer standardized, open-source environments upon which developers can build, test, and validate novel RAN-related agent functionalities. To foster a sustainable and collaborative open-source ecosystem for agentic tools within RAN, we advocate for the creation of standardized guidelines that govern both specification and discovery processes. These guidelines should address not only syntactic compatibility but also execution guarantees, trust, and integration policies.

There are two fundamental dimensions to consider when specifying agentic tools for RAN: (i) Functional Mapping: identifying and mapping key capabilities required across various layers of the RAN protocol stack—ranging from signal processing to configuration and monitoring. (ii) Interaction Scope: defining the scope of agent interactions, including the information elements, service primitives, and control/management entities involved. Any tool integrated into RAN infrastructure must adhere to strict requirements for safety, reusability, and discovery. Without such standardization, siloed or proprietary implementations risk fragmenting the ecosystem and introducing inconsistent or unpredictable behaviors in live workflows. The core standardization requirements for tool on-boarding are summarized in Table 1.

Fine-tuning LLMs for RAN-specific agent applications requires standardization to ensure reliability, efficiency, and adaptability. One widely adopted technique is Retrieval-Augmented Generation (RAG) (Lewis et al., 2020), which enhances LLMs with external knowledge sources to improve task-specific performance. Another method, Low-Rank Adaptation (LoRA) (Hu et al., 2021), offers a computationally efficient approach to fine-tuning while preserving performance. An emerging trend in this domain is Telco-RAG (Bornea et al., 2024), an open-source RAG framework tailored to telecommunications standards. While Telco-RAG shows promise, the complexity of 3GPP (3GPP, 2025) and O-RAN (O-RAN Alliance, 2025) specifications presents significant challenges for LLM integration.

There is an urgent need to define best practices for fine-tuning approaches that yield validatable and reproducible results. Currently, validation mechanisms rely on limited datasets such as TeleQnA (Maatouk et al., 2023), which may not fully reflect real-world telecom environments. Developing richer benchmarking datasets is essential for advancing model reliability and domain adaptation. Given the multi-modal nature of telecom data—including structured tables, logs, textual specifications, and performance metrics—there is growing interest in structured methodologies for multi-modal RAG. Two prevailing approaches exist, and selecting the most suitable paradigm for telecom-specific modalities is critical.

The effectiveness and safety of agent-based implementations in RAN are heavily influenced by the underlying foundation models. Variations across LLMs in terms of latency, throughput, task relevance, and energy efficiency (Laskaridis et al., 2024; Nezami et al., 2024) highlight the urgent need for standardized benchmarking and validation frameworks. To ensure fair, reproducible, and comprehensive evaluation, agent performance must be quantified through structured methodologies, as outlined below:

As 6G-RAN documentation evolves, it is crucial to make specifications accessible not only to human engineers but also to autonomous agents. Current formats are predominantly text-based and optimized for manual interpretation, making them poorly suited for automated reasoning or ingestion (Perez and Vizcaino, 2024). To support agent-based operations in telecom, we propose the standardization of agent-readable specification formats that fulfill the following objectives:

In agent-based systems, a persona defines the agent's role, expertise, and interaction style. While commonly associated with chatbots, recent studies (Mao et al., 2023) show that persona design critically affects orchestration and decision-making in broader applications. Inaccurate or vague personas (Mao et al., 2023) can lead to hallucinations, task misprioritization, or erratic behavior—risks that are unacceptable in RAN, where precision and compliance are vital. To ensure consistency and safety, we propose the standardization of RAN-specific personas along the axes detailed in Box 1.

As next-generation web and handset-based applications evolve, agent-based intelligence will become central to enhancing automation, personalization, and network interaction (Kan et al., 2024). These agents will increasingly manage RAN resources, optimize user experience, and support application logic at the edge. Efficient deployment of such agents hinges on managing computational loads through edge offloading and caching (Zou et al., 2023), particularly given the demands of models like RAG. We propose an investigation into the following key areas to facilitate agent implementation on RAN:

As agent-based systems evolve within telecom networks, seamless integration with external services is essential. A key enabler of such integration is standardized API access, allowing agents to interact with network functions, cloud platforms, and third-party services. The CAMARA project (CAMARA, 2025), under the Linux Foundation, defines and develops APIs for the GSMA Open Gateway framework (GSMA, 2025), which aims to make telecom networks programmable via open interfaces. Integrating CAMARA APIs into agent-based ecosystems can significantly enhance agent capabilities, enabling real-time data access, dynamic service invocation, and scalable interoperability across heterogeneous systems. This integration should focus on:

In RAN implementations, deploying multiple agents across network layers is essential for optimizing performance, automating operations, and enhancing decision-making. These agents—often based on a shared LLM—must collaborate in real time to manage tasks effectively. However, orchestrating and managing a swarm of such agents introduces new complexities that necessitate standardization. AgentOps, the structured approach to agent development, deployment, and monitoring, is critical for integrating these agents into RANOps—the broader operational framework for RAN automation. Key challenges include autonomous coordination, role assignment, and lifecycle management across heterogeneous agents and network layers. The following outlines the primary standardization areas required to support scalable multi-agent orchestration within telecom environments.

To enable robust and intelligent behavior within GenAI-driven RAN architectures, agents must be equipped with standardized reasoning frameworks that combine classical context-aware methods with modern generative techniques. Traditional approaches rooted in information and communication theory (Glover and Grant, 2009; Shannon, 1948) have long supported effective decision-making in telecom systems. Integrating these with GenAI enhances the adaptability and robustness of agents, supporting more dynamic and resilient RAN operations.