Towards Agentic AI Networking in 6G: A Generative Foundation Model-as-Agent Approach

TL;DR

The paper addresses the limitations of data-centric network AI by proposing AgentNet, a specialized AI-native networking framework for agentic AI in 6G and beyond. It introduces a generative foundation model-based implementation with GF-agents across application, physical, and network layers to bootstrap embodied agents and enable knowledge transfer. Through two application scenarios—digital twins-based industrial automation and metaverse-based infotainment—it demonstrates cross-layer task-driven collaboration and reports improvements in CSI prediction accuracy (above 89%) and spectrum efficiency (up to 46% improvement) through coordinated GF-agent actions. The work highlights a path toward scalable, secure, and autonomous AI-enabled networks, while outlining open problems in agent behavior monitoring, multi-agent coordination, and hybrid integration with existing data-oriented networking functions.

Abstract

The promising potential of AI and network convergence in improving networking performance and enabling new service capabilities has recently attracted significant interest. Existing network AI solutions, while powerful, are mainly built based on the close-loop and passive learning framework, resulting in major limitations in autonomous solution finding and dynamic environmental adaptation. Agentic AI has recently been introduced as a promising solution to address the above limitations and pave the way for true generally intelligent and beneficial AI systems. The key idea is to create a networking ecosystem to support a diverse range of autonomous and embodied AI agents in fulfilling their goals. In this paper, we focus on the novel challenges and requirements of agentic AI networking. We propose AgentNet, a novel framework for supporting interaction, collaborative learning, and knowledge transfer among AI agents. We introduce a general architectural framework of AgentNet and then propose a generative foundation model (GFM)-based implementation in which multiple GFM-as-agents have been created as an interactive knowledge-base to bootstrap the development of embodied AI agents according to different task requirements and environmental features. We consider two application scenarios, digital-twin-based industrial automation and metaverse-based infotainment system, to describe how to apply AgentNet for supporting efficient task-driven collaboration and interaction among AI agents.

Figures (4)

• Figure 1: The agent development pipeline (top-right subplot) and the general architectural framework of AgentNet (bottom subplot) including infrastructure (cloud data center, edge servers, UEs, and network infrastructure), environment (physical reality and virtual reality environments), application tasks and goals, agents and agent controller.
• Figure 2: KPI Requirements of two application scenarios, metaverse-based infotainment and digital twins-based industrial automation, on the 6 key performance metrics, measured by model generalization error (Model), non-i.i.d. level of environmental sensing datasets (Environment), knowledge-associated domain specialty (Knowledge), computational and communication resource cost measured by energy consumptions (Resource), task goal relevant inception scores (Goal), and model bias under adversarial attacks (Security), respectively, all in normalized values.
• Figure 3: An implementation of AgentNet based on GF-agents for digital twins-based industrial automation: the predicted network traffic with and without the added robotic arm is shown in the right-top subplot, the spectrogram of physical environment caused by the operation of the added robotic arm is shown in the right-middle subplot, and the final evaluation results about physical-layer and network-layer impacts are listed in the text at the right-bottom subplot.
• Figure 4: An implementation of GF-agent-based AgentNet for metaverse-based infotainment system: the estimated CSIs in 5 uplink and downlink channels are shown in the middle-left and middle-right subplots, respectively, and the estimated bandwidth required with and without adjusting video resolutions for the network layer is shown in the center-top subplot.