Problem

As 6G networks emerge, the focus is shifting from simply increasing throughput to reliably achieving task objectives under constraints of limited bandwidth, latency, and energy. Traditional communication schemes, which prioritize bit-perfect reconstruction, often struggle in new scenarios such as multi-agent collaboration and immersive communication. There is a critical need for systems that can adapt resource scheduling and coding strategies to service objectives and network conditions, leveraging artificial intelligence. While semantic communications (SemCom) and agentic AI offer promising directions, a systematic approach to integrate agentic AI's capabilities (perception, memory, reasoning, and action) into SemCom is essential to move beyond static, single-modal, offline configurations towards dynamic, multimodal, and online optimization.

Method

This paper proposes a unified agentic AI-enhanced Semantic Communications (SemCom) framework that integrates three layers: an application layer, a semantic layer, and a cloud-edge collaborative layer, forming a closed loop of "intent-encode-transmit-decode-act-evaluate." The application layer handles user intent articulation and Quality of Service (QoS) evaluation. The semantic layer performs context-aware semantic encoding and decoding, alongside agent-driven resource orchestration and transport control. This layer incorporates embedded intelligent agents within SemCom components, such as Joint Source-Channel Coding (JSCC) modules, to enable autonomous perception, selective compression, and encoding of task-relevant semantics. Reinforcement Learning (RL) agents are also utilized for resource scheduling, dynamically adjusting parameters like bitrate and bandwidth based on network state and semantic importance. The cloud-edge collaborative layer provides Large Language Model (LLM) and Large Vision Model (LVM) enhancements, offering global management of policies, knowledge, and models through a shared knowledge base (KB).

As a case study, the paper introduces Agentic KB-based JSCC (AKB-JSCC). In this framework, an LLM/LVM agent builds the source KB, serving perception and memory functions by generating and retrieving cross-modal prompt embeddings to enhance the JSCC encoder. Concurrently, an RL agent implements the channel KB, fulfilling reasoning and action roles by adaptively selecting the coding bitrate based on semantic importance and channel state information (CSI). This combined agentic KB aims to achieve efficient and robust SemCom transmission.

Results

The AKB-JSCC framework was evaluated using the MSCOCO 2014 dataset. Experimental results demonstrate that AKB-JSCC achieves superior information reconstruction quality, measured by Peak Signal-to-Noise Ratio (PSNR), across various channel bandwidth ratios (CBR) and signal-to-noise ratio (SNR) settings compared to baseline schemes like NTSCC, JPEG2000+LDPC, and JPEG+LDPC. Specifically, at a CBR of 0.02, AKB-JSCC showed approximately a 9% PSNR gain over NTSCC. The framework also successfully preserves image structures and textures more faithfully, particularly at low CBRs where NTSCC exhibits noticeable blocking artifacts. Furthermore, the entropy and rate preset maps indicate that AKB-JSCC effectively allocates more resources to salient regions (e.g., a train) while dedicating fewer resources to background areas (e.g., the sky), showcasing semantics-aware adaptive resource allocation. An ablation study confirmed that the LVM in the source KB significantly contributes to capturing cross-modal semantic features and strengthening the JSCC codec, even without the channel KB. The study also highlighted SemCom schemes' effectiveness in reducing transmission overhead and mitigating the "cliff effect" common in traditional separated coding schemes.

Implications

This research provides a systematic foundation for integrating agentic AI into semantic communications, offering a pathway toward portable, verifiable, and controllable deployments for 6G networks. The proposed unified framework and its successful application in the AKB-JSCC case study demonstrate the significant potential of combining agentic AI's autonomous capabilities with SemCom's efficiency. The findings suggest that this paradigm can significantly enhance communication robustness and resource efficiency in complex, resource-constrained environments. Looking forward, the authors identify critical research directions for future evolution, including the standardization of protocols, signaling, and APIs to ensure cross-domain consistency and interoperability between agents and SemCom components. Furthermore, establishing security and trust frameworks is vital to address potential issues like data poisoning and information leakage in complex agentic AI-enhanced SemCom systems. Finally, the paper emphasizes the need for unified evaluation metrics and the creation of open testbeds to facilitate reproducible research, comparative baselines, and the seamless transition of agentic SemCom solutions from simulation to real-world deployment across diverse applications such as smart healthcare, satellite positioning, and embodied robotics.