Applicability of Autonomic Networking Ecosystem to 6G AI-Network Integration

AI agents are entities with embedded artificial intelligence to interact multi-modally with applications, end devices and network components. AI agents may exist either in the physical state as embedded HW devices (e.g., robots) or in the virtual state (e.g.,software-implmented applications). With the integration of LLMs, AI agents can understand complex requests, translate them into actionable insights, and orchestrate various services (e.g. communication service, data analyzing service, AI-related services). AI agents play a crucial role in the Telecom domain by enhancing the network efficiency via dynamically optimizing resources, predicting network conditions, making autonomous & intelligent decisions, and facilitating seamless communication among serviced & servicing entities .

With the imminent full unfolding of 6G era, the future world is expected to be full of AI agents. In light of the seeming differentiation between the AI-agent centric and the human-object oriented communication modes, it is highly imperative to define a new network framework that can be tailored to advance the communication among AI-agents, while simultaneously satiating the requirements of existing network entities. This new network framework is named as the AI-agent Network or AAN in the draft. AAN targets at architecting a globally interconnected network to satisfy the on-demand communication, interactions & collaborations with secure and controllable information flow paths for AI-agents in distributed deployment environment, regardless of their instantiation formats (i.e., being in physical or virtual state), capability disparities (being in high-, mid- or low- rank), and/or the hosting devices . Commonly integrated with LLMs, AI-agents demand high computing power and significant energy consumption, which deems the versatility of the specific realization forms of AI agents. For example, an AI-agent can be instantiated as a standalone physical body, or as an intelligent service (in software state) deployed inside the hosting network, or as a cloud-native instance residing in the edge or remote cloud data centers (DCs), or even as hybrid composite entity integrating all the advantages of physical body, hosting network and cloudified deployment.

The versatile instantiation formats of AI-agents may result in three typical architecture and realization models for AAN, namely:

Static intra-domain only AI-agent Communication: A network realization model for the communication among AI-agents residing within a single administrative domain or network. AI-agents in the domain form a network group maintaining the static communication association. AI-agents inside the domain do not communicate with AI-agents outside the domain.
Static inter-domain AI-agent Communication: A network realization model for the communication among AI-agents that reside across multiple administrative domains. AI-agents across these domains form one or more network groups maintaining the static communication associations. Note that in this model, AI-agents in a domain can communicate with AI-agents both inside and outside the domain.
Dynamic multi-domain AI-agent Communication: A network realization model for the communication among AI-agents that may dynamically form a network group to handle a temporarily-generated task. These AI-agents could be in the same or different administrative domains. Once the temporary task is finished, the dynamically-formed network group would be released and the run-time states and session(s) among the involved AI-agents are terminated. This is a dynamic mode versus the previous two static modes.

The variations among AI-agents imply their capabilities (either statically provisioned resources or dynamic runtime loads), the AI related demands, and the data processing tasks vary dramatically. For example, the limited compute power of lightweight terminals, such as smartphones, XR glasses, etc., makes it difficult to handle locally the complex computation tasks with more than billions of parameters. In contrast, if all the complex tasks are delegated to more advanced cloudified instances (residing in a remote data center), then it might impair the real-time responsiveness if the remote instances experience the bursty load of multi-users. Therefore, it is more desirable to consider the seamless collaborations among end devices, networks and (edge or remote) clouds to build a composite AI-agent network, in which AI logics are distributed either at the network edge or within the network, either inside or across domains. In that, the compute power of end devices may realize hierarchical AI reasoning with the help of more powerful network-based entities, which expands the end-side AI services on demand. Further, the network entities can provide more advanced AI services, e.g., Ambient IoT , integrated sensing , etc., to supplement the requirements of (end) AI-agents. The ultimate goal would be the better balance between achieving the intelligence at AI-agents and the lower energy consumption (of course, potentially more advantages). AI-agents offering varied types of services could reside in any domain in a (multi-domain) AI-agent Network. Supposing a complex task is comprised of multiple sub-tasks that need to be handled in a sequence, and every sub-task may be potentially serviced by more than one AI-agent. If these AI-agents are distributed across different domains of an AAN, the service-chaining formed from the (across-domain) AI-agents makes the task processing more challenging. In this scenario, the selection of an optimal AI-agent (among all candidates residing either inside the network or at the edge) for each sub-task may help fulfill the complex task more efficiently. Note that the similar optimization scenarios can be applied to AI-ASA (AI-powered Autonomic Service Agent) in the context of ANIMA.

The 3GPP document TR 23.801 , i.e., the Spec. for the study of the 6G architecture, lists two core research issues (or KIs) for the study of the 6G Network-AI integrated architecture: the research issue concerning AI for 6G Architecture and the other issue on 6G Network for AI. The existing 6G preliminary architecture is not yet mature sufficiently for a unified interactive paradigm between network intelligent entities and distributed edge AI agents, resulting in two core decoupling problems:

AI-for-Network (*): Edge AI capabilities cannot be effectively utilized for network autonomous optimization
Network-for-AI: 6G network resources cannot be dynamically adapted to diverse edge AI service requirements

(*): Note that while both key issues are critical, this document will focus only on the 6G AI-for-Network problem. Regarding this issue, the 6G network relies on built-in AI agents to implement autonomous operation, but cannot actively perceive, screen and utilize the sensing, reasoning and decision capabilities of distributed edge AI-agents. Thus, the ultimate objective is to link the edge AI capabilities (as provided by edge AI-agents) with the in-network closed-loop optimization. This section starts from the categorization of 6G intelligent agents (i.e., AI-agents), and then discusses how the 6G AI-agents may collaborate to strive for the 6G network optimization. This section will also describe some requirements for the potential framework supporting the engagement of (6G) network and edge AI-agents.

There are two core AI-agents defined for 6G architecture:

6G Network Intelligent Agent (6G-NA): The 6G-NA entity can be deployed in core network, RAN and edge network. It undertakes the role of network capability provider and network optimization demander, and is responsible for exposing network resource capabilities and requesting the support of edge AI capabilities.
6G Edge Intelligent Agent (6G-EA): the 6G-EA entity is deployed in distributed mode, residing in end-user devices, edge servers and industrial terminals, covering vision, large model, control and other AI agents. It undertakes the role of edge AI capability provider and network resource demander, and is responsible for providing AI perception reasoning capabilities and putting forward network resource requirements.

These two types of AI-agents require the support of both bidirectional capability perception and dynamic capability negotiation. The bidirectional aspect requires the 6G network AI-agents (i.e., 6G-NA) to perceive the edge AI-agents' (i.e., 6G-EA) capability attributes, and the 6G-EA to acquire the resource capability attributes of the 6G-NA, realizing transparent mutual discovery of heterogeneous intelligent bodies. Further, the 6G-NA and 6G-EA should support the real-time negotiation, matching and adaptive adjustment between 6G network resource supply and edge AI business demand, breaking static resource binding limitations.

The 3GPP 23.801 defines three standardized network-edge AI-agent interaction modes, covering all core 6G AI-network collaborative scenarios. Namely, the three modes are the AI for 6G network optimization mode, the 6G network-empowered edge AI service mode, and the dual mode via the bidirectional collaboration of 6G-NA and 6G-EA. The draft focuses on the AI for 6G network scenario, in which the edge AI-agents (6G-EAs) empowers network autonomous operation and closed-loop optimization. The distributed-deployed 6G-EA actively exposes its own learning, reasoning and state prediction capabilities, while the in-network 6G-NA screens eligible edge (AI) agents through negotiation, obtains edge real-time context and AI analysis results, and then guides the network adjustment, mobility management, self-protection and self-repair. This type of interaction mode revolves around addressing the problem of insufficient network perception capability, which may realize the AI-driven autonomic network operation & optimization.

The 6G preliminary architecture, as described in the TR 23.801 , lacks a unified interactive paradigm between the 6G network intelligent entities (i.e., 6G-NA) and distributed edge AI agents (i.e., 6G-EA). This results in the critical issue that this draft pursues to tackle, i.e., the capabilities of 6G-EA cannot be effectively utilized for the autonomous optimization of 6G network. Therefore, a potential framework covering both 6G-EA, 6G-NA and their interactive behaviors needs to be studied. The framework should take 6G-EA's & 6G-NA's capabilities of mutual discovery and dynamic negotiation as the core, define standardized interactive scenarios and application modes for 6G NA-EA collaboration, and put forward the necessary system-level requirements to support the deep integration of AI with the 6G networks.

As defined in , an Autonomic Service Agent or ASA is an agent implemented on an autonomic node that instantiates an autonomic function, either in part (in the case of a distributed function) or whole. An ASA makes use of the features provided by the Autonomic Networking Infrastructure or ANI , to achieve its own goals, usually including interactions with other ASAs via the signalling exchanges thru ANIMA defined protocol, e.g. GRASP . AI-ASA's are AI-powered ASAs with intelligent management functions (AFs). As defined in the 3GPP TR 23.801 , both 6G-NAs and 6G-EAs are automated, intelligent entities that can, via the interactions with their environment, learn & analyze knowledge, make decisions, and then execute tasks to achieve specific goals. They can operate intelligently & completely on their own or collaborate with other 6G- NAs or EAs in the mobile domain. Empowered by the intelligent management functions, both 6G- NAs and EAs can be categorized as AI-ASAs as in the context of ANIMA.

The 3GPP SA2 WG is currently discussing the evolution of the 6G system architecture. Out of the 6G CP, UP and management planes, the 6G UP is of significance to the theme of the draft. One of the 6G-UP objectives targets at enhancing the 5G-like UP architecture to achieve flexiblity, scalability and resilience. For this matter, the TR 23.801 has proposed to introduce new 6G Network Functions or NFs, spanning both UP and CP, as well as the concept of 'set of UP NFs'. The shows briefly the 6G UP architecture. Both S-UPF and A-UPF are newly proposed 6G NFs, with the S-UPF targeting at supporting basic UP functionalities and the A-UPF for more advanced features. The DMM draft on 5G & 6G MUP evolution provides more details .

6G UP Architecture w/ AI-ASA (6G NA & EA) 6G-EA : (AI-ASA) : (AI-ASA) : / \ : +----+ : / \ : |E-DN| Mobile : +---+ +-------+ +-------+ : /+----+ User----+RAN+---+ S-UPF |---->>| A-UPF |---\ : / (UE) : +---+ +-------+ +-------+ +-->>--+ : (AN) \ (AN) / : \ : |<....(ACP)....>| / : \ : \ / : +----+ : \ ---->>---- / : |C-DN| : : +----+ ........................................ ]]>

The describes four components that constitute together the base elements of the autonomic networking ecosystem. Namely, they are autonomic service agents (ASA), autonomic network infrastructure (ANI), autonomic control plane (ACP), and the GeneRic Autonomic Signaling Protocol (GRASP). The exhibits the mappings and relationship between components of autonomic ecosystem and various entities of 6G. Here the E-DN indicates the edge data network for the deployment of 6G-EA. Both S-UPF and A-UPF could take on the role of autonomic nodes to employ autonomic functions.

AI-powered ASA: Both 6G intelligent agents, e.g., 6G NAs and EAs, are AI-powered agents or AI-ASAs if autonomic functions (AFs) have materialized, whose bidirectional collaborations target for network optimization. Both may add the capabilities to implement autonomic functions. The 6G-NA can be implmented on S/A-UPF in the 6GS, while the 6G-EA can be employed in the E-DN, as demonstrated in the figure.
Autonomic Networking Infrastructure (ANI): Contain common control and management functions, such as messaging, service discovery, negotiation, Intent distribution, self-monitoring, and diagnostics, etc., for the operations of 6G- NAs and EAs.
Autonomic Control Plane (ACP): Targetting for the control-plane operations among autonomic nodes (i.e., S/A-UPFs) that host AI-ASAs, i.e., 6G NA and 6G EA.
GRASP: For information exchanges among AI-ASAs, i.e., 6G- NA and EA in the context of 6G, which negotiate & signal AI-generated policies. Note that the extreme dynamics of some AAN realization models, as shown in the , post extra challenges to the signalling protocols that may require the enhancement to the GRASP.

There is no security concern.

There is no IANA requirement.