Problem

Current 5G systems face challenges in coverage, cost, energy efficiency, and customization for various industries. The advent of agentic AI, with its capabilities in multi-modal sensing, complex task coordination, and self-optimization, is driving the evolution towards agent-based communication networks, especially in the context of 6G. Traditional bit-level transmission struggles to support these demands due to high data volumes, low reliability, and excessive resource consumption. Semantic Communication (SemCom) offers a transformative solution by focusing on task-oriented efficiency, enhanced reliability, and dynamic resource adaptation. However, there is a lack of comprehensive reviews that systematically explore the technological evolution of semantics within agent communication networks.

Method

This paper addresses the gap by systematically exploring the role of semantics in agent communication networks. The authors propose a novel architecture comprising three wireless agent network layers (intention extraction and understanding, semantic encoding and processing, distributed autonomy and collaboration), four AI agent entities (embodied, communication, network, and application agents), and four operational stages of semantic-enhanced agentic AI systems (perception, memory, reasoning, and action). These components are designed to interact in a closed-loop system. Based on this architecture, the paper provides a comprehensive review of state-of-the-art technologies that enhance agent communication networks through semantics, analyzing advancements across each layer, entity, and stage.

Results

The comprehensive review based on the proposed architecture highlights numerous advancements. In the intention extraction and understanding layer, methods include evidence-based, opponent modeling-based, and mind modeling-based intent inference. For the semantic encoding and processing layer, key technologies involve semantic-based coding (JSCC, generative coding), semantic-based beam management (visual and channel assisted), semantic-based CSI feedback (reconstruction, knowledge-driven, optimization), semantic-based HARQ (similarity, feature, adaptive), Age of Semantic Information (reconstruction, task-oriented), and semantic Knowledge Bases (construction, deployment). The distributed autonomy and collaboration layer leverages distributed access (MDMA, semantic fusion), knowledge collaboration (federated semantic learning, multi-agent alignment, semantic relaying, collaborative inference), and resource scheduling (query-semantic, semantic importance-aware, adaptive). Furthermore, semantics enhance the agentic AI stages: perception (semantic feature extraction, task-oriented sensing, object grounding/tracking, embodied environment understanding), memory (hierarchical structures, retrieval and reasoning, evolution/update, cognitive augmentation), reasoning (CoT, KG-augmented, retrieval-augmented, tree-structured multi-path, neuro-symbolic), and action (semantic tool acquisition, reasoning-action interleaving, multi-agent collaboration, self-correction, reinforcement-based feedback). The paper also identifies and discusses representative projects and implementations for each of the four AI agent entities.

Implications

Despite significant progress, several fundamental challenges remain for semantic-based agent communication networks. A unified mathematical foundation is needed to define and quantify semantic information and its impact on decision-making, as well as to characterize performance boundaries for joint communication, computation, and intelligence optimization. Managing semantic Knowledge Bases presents challenges in fast alignment, efficient dynamic updates, preventing interference between old and new knowledge, and designing lightweight representations for resource-constrained agents. Security and privacy protection are critical, as semantic attacks could alter meaning with minimal data tampering or propagate hallucinated information, requiring research into encrypted semantics, verifiable authenticity (e.g., blockchain), trusted execution environments, and interpretable models. Finally, promoting standardization and industry adoption across the entire value chain is crucial, including defining cross-domain semantic translation standards, establishing open-source platforms, and developing hybrid bit/semantic communication technologies.