The diverse and heterogeneous terminal artificial intelligence (AI) agents and network-element AI agents are flourishing in a flywheel-like manner. The new characteristics of their capabilities and behaviors will reshape the service paradigm and traffic logic of future mobile information networks. This article first elaborates on the dynamics of the intertwined and integrated development of the AI agent/robot industry and the wireless communication industry. Then, based on an analysis of the new capabilities and behavioral characteristics of terminal AI agents and network-element AI agents, the article deduces a new design paradigm for future AI agent-centric (AA-Centric) networks, which includes seven core features: intent-driven, proactive service, distributed collaboration, efficient customization, deterministic guarantee, online evolution, and infinite generation. Guided by this new paradigm, the new architecture of 6G networks is further deduced, and the key supporting technologies are expounded. Finally, it is concluded that AI agents play a crucial role in driving the innovation of future network architectures and the ultimate expansion of capabilities.

As 6G approaches, the proliferation of large language models (LLMs) and embodied intelligence is driving a paradigm shift from the Internet of Things (IoT) to the Internet of Agents (IoA). However, traditional network architectures, designed for content-agnostic data transmission, struggle to accommodate the bursty, reasoning-driven traffic patterns and rigorous multimodal synchronization requirements of autonomous agents. This paper surveys the AI-agent communication network (ACN), aiming to bridge the gap between static network resources and dynamic agent tasks. We analyze the evolution from bit-oriented transmission to agentic syntax protocols, which enable intent-based signaling and semantic compression. Furthermore, we explore mechanisms for multi-agent collaborative consensus and distributed decision-making under the constraints of unstable wireless environments. We critically focus on task-driven dynamic networking, examining how integrated sensing, communication, and computing (ISCC) and network-embedded agents (NEA) facilitate the real-time generation of task graphs and intent-aware traffic scheduling. To synthesize these technologies, we propose a reference framework, the Deep-Agentic Network Architecture (DAN-Arch), which vertically integrates physical-layer sensing with application-layer reasoning flows. Finally, open challenges regarding energy efficiency, cross-domain governance, and 3GPP standardization pathways are discussed to guide future research towards a fully agent-native 6G ecosystem.

With the emergence of AI-agent communication networks (ACN) in the 6G era, the efficient training of agents for complex reasoning tasks has become a critical capability for scalable ACN deployment. As a representative complex reasoning task, retrieval-augmented multi-hop question answering (e.g., agentic retrieval-augmented generation) requires agents to perform multi-step reasoning through reflection, planning, and tool-use mechanisms. However, reinforcement learning training still faces reward sparsity and sample efficiency challenges, limiting agents’ rapid evolution and adaptability. We propose dynamic batch-aware advantage weighting (DB-AW), integrating two core components at the batch level: the difficulty-aware weighting component dynamically amplifies positive advantages based on long-term success rates, directing learning toward learnable yet challenging samples; and the batch filtering component removes zero-variance groups, ensuring each update contains non-zero gradient signals. Experiments show that DB-AW achieves 18%, 17%, and 15% relative improvements on Qwen2.5-7B, Qwen2.5-3B, and LLaMA3.2-3B, respectively, while improving the effective update rate from 68% to 100%, significantly reducing agent training costs. As a lightweight and reusable algorithmic module, DB-AW can be readily integrated into methods such as group relative policy optimization (GRPO), providing a practical pathway for efficient training of complex reasoning agents in ACN.

With the rapid advancement of generative artificial intelligence (AI) and large language model (LLM) technologies, AI agents are gradually becoming the core service units in networks, and their communication mode is evolving from local collaboration to wide-area interconnection. The construction of the Internet of Agents (IoA) faces multiple challenges, such as identity management, dynamic networking, and semantic routing, which urgently requires the design of a network protocol system that adapts to its new traffic characteristics and collaboration needs. Based on the application scenarios of agent communication, this paper systematically analyzes the management, control, and routing requirements that multi-agent collaboration imposes on IP networks, proposes a three-layer functional architecture for the IoA, and designs a protocol suite covering management, control, and routing around key issues such as agent registration and identification, service discovery, capability sensing, and cross-domain traffic assurance. By extending existing Internet protocols and introducing a semantically aware routing mechanism, this paper provides a scalable, efficient, and secure approach to implementing a protocol for end-to-end agent collaboration, thereby contributing to the construction of an open, large-scale agent collaboration ecosystem.

The convergence of artificial intelligence (AI) with the physical world is reshaping the future of intelligent systems through real-time perception, interaction, and control within physical environments. To support this new paradigm, 6G networks are envisioned as critical enablers, offering ultra-low latency, high reliability, and service-aware intelligence to facilitate seamless human-machine collaboration. This paper proposes a functional framework that integrates Agentic AI into the 6G architecture, introducing the concept of Agentic AI-Enabled 6G Network Services (AA6NS). In this framework, user intents are translated and processed across the application layer, core network (CN), and radio access network (RAN), where Agentic AI dynamically manages task-level quality of service/quality of experience (QoS/QoE), orchestrates multi-device service groups, and enables real-time network adaptation. The proposed architecture with the new 6G techniques establishes a foundation for future physical AI applications across domains such as autonomous mobility, smart manufacturing, and remote robotics.

Empowered by advances in large language models, the growing integration of autonomous agents into industrial and daily-life sectors is turning them into new networking entities. Such agent-oriented networking features high interaction frequencies and emergent task-driven structures, necessitating strong network policy consistency and reliability within dynamic environments. To address these challenges, we propose a network control system that integrates Intent-Driven Network (IDN) into Heterogeneous Agent-Oriented Networking (HaoNet). IDN focuses on high-level task intents and provides flexible reconfiguration and adaptive optimization, thereby enhancing the effectiveness of agent-oriented networking. In this paper, we first summarize three key features of HaoNet: task-driven operation, distributed collaboration, and closed-loop intelligence. Furthermore, we propose a comprehensive system architecture, which includes the application layer, the intent layer, and the infrastructure layer, and investigate the associated key technologies. Finally, typical application scenarios are presented to demonstrate the practical value of the proposed system in enabling robust agent-oriented networking control.

The emergence of multi-agent systems (MAS) based on large language models (LLMs) has enabled autonomous collaboration on complex, goal-oriented tasks. However, effective interaction is frequently hindered by semantic drift, a phenomenon where heterogeneous agents assign conflicting meanings to shared terminology due to differing internal prompts or domain knowledge. Existing communication paradigms either rely on unconstrained natural language, which suffers from structural vagueness, or rigid symbolic schemas that fail to adapt to emergent concepts. To address this gap, we propose DOA, a novel dynamic ontology alignment framework that serves as a semantic mediation layer for MAS. DOA integrates a proactive semantic prober to detect conceptual mismatches and a neuro-symbolic aligner that reconciles local semantic structures in real time. By grounding fluid natural language dialogues in an evolving shared ontology, our framework ensures deterministic mutual understanding over long-horizon tasks. Empirical evaluations in cross-domain supply chain and healthcare coordination scenarios demonstrate that DOA improves task success rates by an average of 31.5% and reduces communication overhead (token consumption) by 50% compared to state-of-the-art baselines. Our results provide a robust and scalable foundation for semantic consistency in next-generation industrial-grade AI systems.

With the sustained growth of live video streaming, the demand for high-quality video services for mobile users across diverse network environments is increasing rapidly. In this paper, we define the network fluctuation characteristics in different environments as network modality. To comprehensively investigate network modality across different environments, we construct a network modality dataset by collecting multi-dimensional network metrics from various real-world scenarios and suggest that network modality exhibits separability. Therefore, network modality recognition, which aims to distinguish the scenarios where a user is located based on network modality sequences, is feasible and can be formulated as a multivariate time series (MTS) classification problem. To address this problem, we propose a novel neural network (NN)-based classification model called ACTap. Specifically, the model first integrates a two-stage attention (TSA) mechanism and a convolutional neural network (CNN) to extract features from network modality sequences. Then, it filters out noisy feature representations to learn discriminative class prototypes, and finally recognizes network modality based on the distance between their feature representations and class prototypes. Experimental results validate the separability of network modality and show that ACTap outperforms four benchmark models in terms of classification accuracy on the network modality dataset.

Holographic multiple-input multiple-output (HMIMO) systems deploy ultra-dense antennas in confined spaces, yet spatial degrees of freedom (DoF) fail to scale with element count. Only by harnessing evanescent waves can the potential gains of ultra-dense arrays be unlocked. However, in practical scenarios, antenna apertures are typically too small relative to communication distances to generate significant near-field effects for capturing these evanescent waves. This paper proposes a method to alter the dispersion relation via locally resonant metamaterials (LRM), thereby enabling the radiation of near-field evanescent wave components into the far field. This approach leverages the DoF gains offered by ultra-dense elements within a confined aperture. Full-wave simulations validate the effectiveness of the evanescent-propagating wave conversion method, demonstrating an increase in the spatial DoF radiated into the far field by the HMIMO system even with a limited aperture.

With the rapid advancement of 5G network technology, cyber security threats are intensifying. The fifth-generation mobile communication for railways (5G-R), as a product of the deep integration of the railway industry with 5G technology, plays a crucial role in ensuring the operational safety of railways. Cyber security assessment is one of the key components in safeguarding network security, and attack graphs are one of the mainstream models for cyber security assessment methods. This paper focuses on optimizing the method for assessing the security of the 5G-R core network based on attack graphs. By integrating the business security requirements of the 5G-R core network, this paper optimizes the parameters for the conventional vulnerability assessment methods and the influencing indicators of atomic attack success probability. It also constructs a formula for calculating the atomic attack success probability based on the fuzzy analytic hierarchy process (AHP). By setting up a simulated environment of the 5G-R core network within a network target field, this paper conducts the feasibility verification of the assessment method for 5G-R core cyber security. Comparative experiments with the conventional common vulnerability scoring system (CVSS) and methods used in existing literature prove the validity and superiority of the proposed model.