Motivations and Problem Statement of Agentic AI for network management
draft-hong-nmrg-agenticai-ps-00
This document is an Internet-Draft (I-D).
Anyone may submit an I-D to the IETF.
This I-D is not endorsed by the IETF and has no formal standing in the
IETF standards process.
| Document | Type | Active Internet-Draft (individual) | |
|---|---|---|---|
| Authors | Yong-Geun Hong , Joo-Sang Youn , Qin Wu , Benoît Claise | ||
| Last updated | 2025-10-20 | ||
| RFC stream | (None) | ||
| Intended RFC status | (None) | ||
| Formats | |||
| Stream | Stream state | (No stream defined) | |
| Consensus boilerplate | Unknown | ||
| RFC Editor Note | (None) | ||
| IESG | IESG state | I-D Exists | |
| Telechat date | (None) | ||
| Responsible AD | (None) | ||
| Send notices to | (None) |
draft-hong-nmrg-agenticai-ps-00
Network Management Y. Hong
Internet-Draft Daejeon University
Intended status: Informational J. Youn
Expires: 23 April 2026 DONG-EUI University
Q. Wu
Huawei
B. Claise
Everything OPS
20 October 2025
Motivations and Problem Statement of Agentic AI for network management
draft-hong-nmrg-agenticai-ps-00
Abstract
This document outlines the key objectives of introducing Agentic AI
to the field of network management and highlights the fundamental
issues with existing technologies that must be addressed to achieve
these goals. It emphasizes the necessity for relevant groups within
the IETF/IRTF and presents the core technological areas requiring
standardization. The aim of Agentic AI is to facilitate a paradigm
shift in which multiple autonomous AI agents collaborate to fully
automate network operation, management and security.
Discussion Venues
This note is to be removed before publishing as an RFC.
Discussion of this document takes place on the Network Management
Research Group mailing list (nmrg@irtf.org), which is archived at
https://mailarchive.ietf.org/arch/browse/nmrg.
Source for this draft and an issue tracker can be found at
https://github.com/billwuqin/agentic-ai-ps.
Status of This Memo
This Internet-Draft is submitted in full conformance with the
provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF). Note that other groups may also distribute
working documents as Internet-Drafts. The list of current Internet-
Drafts is at https://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."
This Internet-Draft will expire on 23 April 2026.
Copyright Notice
Copyright (c) 2025 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents (https://trustee.ietf.org/
license-info) in effect on the date of publication of this document.
Please review these documents carefully, as they describe your rights
and restrictions with respect to this document.
Table of Contents
1. Introduction
2. Conventions and Definitions
3. Agentic AI for Network mMnagement
3.1. Role of Agentic AI in Network Operations
3.2. Operation of Agentic AI for Network Management
3.2.1. Intelligence core
3.2.2. Execution & Interaction
4. Problem Statement of Existing Techniques for Agentic AI
4.1. Architectural Bottlenecks and the Failure of Centralization
4.2. Absence of agent-to-agent (A2A) Semantic Interoperability
4.3. Lack of Dynamic Trust and Accountability in Autonomous
Behavior
4.4. Real-time Data Validity and Resilience Issues
4.5. Problems with the Existing IBN System: Rigidity of the
Intent Translation Engine (ITE)
4.6. ANIMA ASA's Problem: Cognitive Simplicity
5. Objectives of Agentic AI for Operations & Management
5.1. Objective 1 - Autonomous Network Operations & Management
5.2. Objective 2 - Intelligent & Dynamic Resource Orchestration
5.3. Objective 3 - Predictive & Adaptive Network Security
5.4. Objective 4 - Enabling Novel Network Service Models
5.5. Objective 5 - Autonomous, High-Fidelity & Action-Aware
Network Measurement
6. Use cases of Agentic AI for Operations & Management
6.1. Intent Based Service Delivery
6.2. Cross-layer and Cross-domain Multi-Agent communication for
Complaint handling
6.3. AI Agent Driven Network Management
7. Security Considerations
8. IANA Considerations
9. References
9.1. Normative References
9.2. Informative References
Acknowledgments
Authors' Addresses
1. Introduction
The explosive growth of digital services and the increasing
complexity of networks in 5G and future 6G environments demand real-
time responsiveness, high efficiency and the ability to make
autonomous decisions on a large scale from operational environments.
To overcome the limitations of existing static automation methods and
human-led Intent-Based Networking (IBN), a new Agentic AI-based
paradigm is required. This involves introducing autonomous software
entities that can interpret information, make decisions, perform
meaningful autonomous actions and adjust plans in response to
changing circumstances.
Unlike traditional automation, which relies on pre-programmed rules,
agentic AI uses autonomous decision-making capabilities to handle
large-scale network activities and customer requests swiftly and
accurately. These agents perform tasks such as network traffic
management, fault resolution, and customer interaction support,
continuously executing responses that previously required manual
human review or escalation.
Agentic AI uses large language models (LLMs) to encompass a wide
variety of capabilities, such as reasoning, problem-solving,
interacting with external environments and performing actions, which
extend far beyond natural language processing. It can decompose
tasks, breaking down complex objectives into specific tasks and
subtasks to achieve them. This cognitive capacity enables a
persistent cognitive cycle (observation, inference, action),
continuously aligning network operations with high-level business
intent.
When such autonomous agents are widely deployed across the
communications and network domains, standardized protocols are
essential to ensure interoperability and security between different
vendor platforms and network domains. The collaborative nature of
agent-based AI systems (multi-agent systems, or MAS) means that
standardized agent-to-agent protocols (A2A protocols) must be defined
to prevent silos forming within the system and to facilitate
discovery, understanding and collaboration between agents.
2. Conventions and Definitions
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
"OPTIONAL" in this document are to be interpreted as described in
BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all
capitals, as shown here.
3. Agentic AI for Network mMnagement
3.1. Role of Agentic AI in Network Operations
The complexity of network management and network operations are
increasing exponentially, due to the increased size of networks and
the increased frequency of change, for for the new 5G and future 6G
services. This makes it increasingly difficult for existing
automation techniques to meet the requirements for operational
efficiency and service quality. Consequently, Agentic AI is an
essential technological advancement for the realization of autonomous
networks.
Agentic AI refers to intelligent systems that can act autonomously to
achieve specific business objectives with minimal human supervision.
These systems can reason through multi-step problems and adjust their
actions in real time. Unlike passive traditional AI systems that
respond only to direct commands, Agentic AI is an active system
operating within an autonomous, closed-loop framework. This
framework enables the system to perceive its environment, reason,
plan a sequence of actions and execute them using various tools and
APIs. This autonomy enables it to perform complex, multi-step
processes such as software development, data analysis and network
management.
The aim of autonomous networks is to leverage the capabilities of
Agentic AI in order to transition operations and maintenance from
static, human-managed, rule-based automation to dynamic, intent-based
automation that is governed by humans. The ultimate goal is to
reduce management costs and complexity, enabling rapid business
optimization at unprecedented levels.
The primary objective of Agentic AI is to enable autonomous decision-
making and the resolution of complex, multi-domain tasks. This is
crucial in bringing operations closer to the level of autonomy that
Agentic AI aims to achieve, by facilitating cross-domain
collaboration. To achieve this, Agentic AI must align network
capabilities with strategic business priorities, such as improving
customer experience and reducing operational costs. This involves
translating comprehensive business intent into localized, actionable
network configuration plans.
Agentic AI optimizes resource allocation based on real-time demand
and business objectives, enabling smarter resource and energy usage.
In architecture research for 6G, for example, the application of
constrained agentic AI techniques focused on energy efficiency and
secure real-time learning for dynamic resource allocation has been
identified as a key objective [Agentic-AI-Wireless].
The Autonomic Networking Integrated Model and Approach (ANIMA)
Working Group of the IETF developed the Autonomic Service Agent (ASA)
for autonomic networking. [RFC7575] defines the ASA as An agent
implemented on an autonomic node that implements an autonomic
function, either in part (in the case of a distributed function) or
whole [RFC7575]. In other words, the ASA is a core component of
ANIMA: a software module that performs autonomic functions on network
nodes. The ANIMA Working Group is defining design guidelines,
lifecycle management, authorization and coordination standards for
the ASA [ANIMA].
IETF’ AI Preferences (AIPREF) Working Group is focused on
standardizing a common vocabulary and mechanism through which users
and systems can express their preferences regarding the use of their
content in the development, training, deployment and use of AI models
[AIPREF].
3.2. Operation of Agentic AI for Network Management
The principal components of agentic AI can be broadly divided into
the intelligence core and the execution tool domain.
3.2.1. Intelligence core
The intelligence core is responsible for an agent's decision-making
and problem-solving capabilities. Large language models (LLMs) or
specialized AI models form the basis of this core. Reasoning Engine/
LLM: This constitutes the core of the agent's brain. It understands
abstract objectives (intent) received from users or higher-level
systems, creates step-by-step plans (plan) to achieve them, evaluates
the outcomes of execution (reflection) and uses logical reasoning to
modify plans or determine subsequent steps.
Memory is the data repository that agents learn from and refer to.
* Short-term memory: It stores the context of the current task and
recent execution results.
* Long-term memory: It stores persistent information such as
previously successful solutions, general knowledge and network
architecture guidelines.
The tool Orchestrator manages the list of external tools (APIs,
functions) available for agents to use. During the planning phase,
it determines which tool is most appropriate and, during the
execution phase, it is responsible for calling the tool and
accurately configuring the necessary parameters.
3.2.2. Execution & Interaction
These components enable the agent to communicate with and make
changes to the external environment (i.e. the network or system).
* Tool set/capability: A collection of all the external interfaces
that an agent uses to perform tasks within a network environment.
* Execution environment: A sandbox environment in which code
generated according to the plan is executed safely, and external
tools are invoked.
* Sensing/observation mechanism: The channel through which the agent
verifies execution results and collects the current environmental
state. This involves more than just invoking tools; it
continuously draws network events, sensor data and similar inputs
into a feedback loop.
+----------------------------------------------------------------+
| |
| +--------------------------+ |
| | 1.GOAL / INTENT(Input) | |
| +------------------|-------+ |
| v |
| +---------+----------------+ |
| | 2.AI AGENT(Brain) | |
| | (LLM/Reasoning Engine) | |
| +---|-----------------|----+ |
| | Memory/Context | | |
| +-----------------+ | |
| ^ v |
| +---------|----------+ +------|------------+ |
| | 4.REFLECT(Compare) | | 3.PLAN(Sequence) | |
| | (Evaluate Outcome) |<---| (Action Breakdown)| |
| +---------|----------+ +------|------------+ |
| | v |
| | +------------------+----------------------+ |
| +---| 5.EXECUTE(Action) via Tool Orchestrator | |
| +------------------|----------------------+ |
| v |
| +---------------------+-----------------+ |
| | 6.TOOL USE(API Calls & Configuration) | |
| | (RESTCONF, Monitoring, etc.) | |
| +-----------|-------------------|-------+ |
| v ^ |
| +-----------+-------------------+--------+ |
| | 7.NETWORK ENVIRONMENT (The World) | |
| | (Apply Changes & Sense/Observe State) | |
| +----------------------------------------+ |
| |
+----------------------------------------------------------------+
Figure 1: Execution & Interaction
4. Problem Statement of Existing Techniques for Agentic AI
4.1. Architectural Bottlenecks and the Failure of Centralization
Existing AI and automation systems have often relied on centralized
infrastructure for data aggregation and heavy computing. However,
these centralized models cannot handle the volume, velocity, and
distributed nature of Agentic AI workloads. Centralized AI systems
are constrained by central infrastructure, resulting in high latency
due to round-trip times to the cloud. Such latency is unacceptable
for real-time applications such as self-healing and 5G slicing
management. There is also the issue that the central server becomes
a bottleneck, limiting scalability. The inherent limitations of such
centralized models (single point of failure (SPoF), latency)
inevitably drive Agentic AI architectures towards a distributed mesh
form. This leverages local processing at the edge for low latency
and fault tolerance, requiring the standardization of distributed
control and communication mechanisms that transcend conventional
centralized SDN/management models.
4.2. Absence of agent-to-agent (A2A) Semantic Interoperability
Agentic systems are often built by different vendors using various
frameworks, leading to fragmented and siloed system operations.
Complex network management tasks require the decomposition of work
and collaboration between specialized agents. Without standardized
agent-to-agent (A2A) protocols, bespoke connectors become necessary
to connect these fragmented systems, slowing down development and
integration speeds.
Standardization must define consistent payloads and interfaces that
support real-time interactions between systems, enabling agents to
discover, understand, and collaborate with one another regardless of
their underlying implementations.
4.3. Lack of Dynamic Trust and Accountability in Autonomous Behavior
The introduction of AI agents as autonomous entities performing
actions at machine speed presents significant security and governance
challenges. Traditional identity and access management (IAM) focuses
on human users or predefined roles. However, autonomous agents
operate with dynamic intent, require context-aware access, and must
maintain provable accountability for every action they perform.
Without a robust Zero Trust framework specifically designed for non-
human autonomous entities, there is a risk of catastrophic security
breaches or manipulation where autonomous systems could outpace human
control capabilities.
4.4. Real-time Data Validity and Resilience Issues
The decision-making of AI agents is determined by the quality of the
data they receive. In a network environment, data quality is of
paramount importance. Incomplete, delayed, semantic-less, context-
less, or corrupted data feeds can lead to severe operational or
financial losses when agents take autonomous actions (e.g., traffic
rerouting, forced execution of financial transactions). Therefore,
it must extend beyond the current focus on bandwidth and speed to
include quality verification of the data agents rely upon and
resilience of service paths. This is essential to meet the
requirements of continuously operating intelligent agents.
4.5. Problems with the Existing IBN System: Rigidity of the Intent
Translation Engine (ITE)
Existing IBN systems rely on the Intent Translation Engine (ITE) or
the Intent-Based System (IBS) spatial functionality to bridge the gap
between the business intent and the network operational
infrastructure. This translation is typically driven by predefined
data models such as YANG models and lacks the necessary adaptive
flexibility when unforeseen conditions arise. IBN fundamentally
shifts operational modes to a dynamic intent-based approach, yet
retains the inherent limitation that control remains under human
oversight. Agentic AI minimises or eliminates human intervention in
this cognitive loop through LLM-based reasoning and planning
capabilities, refining the IBN closed loop by integrating continuous
reasoning and conflict resolution capabilities into the cognitive
layer. These capabilities represent what was lacking in the
classical IBN definition and form the core technical objective.
4.6. ANIMA ASA's Problem: Cognitive Simplicity
ANIMA's ASAs are typically designed for specific, localized
autonomous functions (e.g., prefix management, bootstrapping). They
rely heavily on predefined policy structures and lack the complex
reasoning, planning, or self-reflection capabilities characteristic
of Agentic AI (LLM-based task decomposition). ANIMA's ASA is
conceptually a precursor to Agentic AI, but lacks a cognitive core
(LLM/inference engine). Agentic AI introduces LLM-based planning and
tool-use capabilities that require complex, semantic negotiation
(A2A) beyond simple information exchange (GeneRic Autonomic Signaling
Protocol; GRASP), demonstrating the necessity for a dedicated
protocol layer that extends beyond the existing ANIMA framework.
5. Objectives of Agentic AI for Operations & Management
5.1. Objective 1 - Autonomous Network Operations & Management
Beyond minimizing human intervention, it must implement a Autonomous
Driving Netowrk (defined in TMF) that autonomously recognises,
diagnoses, infers, and resolves issues even in unpredictable
situations.
Key Features:
* Predictive & Proactive Fault Management: AI agents learn traffic
patterns, logs, and performance metrics in real time to identify
potential causes before faults occur. The network autonomously
reroutes traffic or reallocates resources to prevent service
interruptions at source.
* Intelligent Root Cause Analysis: In complex, intertwined fault
scenarios, multiple agents collaborate to synthesize distributed
data. They deduce the root cause as a "problem of correlations"
rather than a single point of failure and propose solutions.
* Autonomous Configuration & Optimization: AI agents comprehend
high-level objectives such as ‘optimize user experience’ and
autonomously configure and continuously fine-tune routing
protocols, QoS policies, security rules, and other elements to
achieve them.
5.2. Objective 2 - Intelligent & Dynamic Resource Orchestration
To address unpredictable traffic demands such as 6G, holographic
communications, and large-scale IoT, network resources (computing,
storage, bandwidth) are allocated and coordinated in real time and
proactively.
Key Features:
* Dynamic Network Slicing: AI agents recognize application
requirements (latency, bandwidth, etc.) in real time, instantly
creating, scaling, and downsizing customized network slices per
user or service.
* Cross-Domain Resource Negotiation: AI agents distributed across
networks of different telecommunications or cloud providers
negotiate in real time to dynamically secure optimal resources,
ensuring end-to-end quality for global services.
* Edge Computing Resource Optimization: By predicting edge node load
and user mobility, AI agents dynamically reallocate workloads to
optimal edge nodes while ensuring service continuity.
5.3. Objective 3 - Predictive & Adaptive Network Security
Beyond defending against known attack patterns, AI agents
autonomously detect unknown zero-day attacks or advanced persistent
threats (APTs) and reconfigure defence systems in real time.
Key Features:
* Autonomous threat hunting and response: Security agents
continuously detect minute anomalies across the entire network.
If an anomaly is deemed a threat, they respond immediately by
taking action such as isolating infected nodes or blocking attack
traffic, all without human intervention.
* Dynamic Defense Posture: AI agents dynamically modify firewall
policies, access control lists (ACLs), and traffic filtering rules
in real time based on attack type and intensity, thereby
minimizing the attack surface.
5.4. Objective 4 - Enabling Novel Network Service Models
By transforming the network itself into a single, vast distributed AI
platform, it enables new communication services and business models
that were previously impossible.
Key Features:
* Intent-driven Service Creation: When a user requests in natural
language, 'I want to play a lag-free VR game with my friends,' an
AI agent interprets this and provides a Network-as-a-Service that
instantly allocates the necessary resources (such as network
slices and edge servers).
* Semantic Communication: Communication focuses on the “meaning” or
“purpose” conveyed by data rather than the bits themselves,
enabling ultra-efficient communication that achieves maximum
effect with minimal data transmission.
5.5. Objective 5 - Autonomous, High-Fidelity & Action-Aware Network
Measurement
To turn raw network telemetry into trustworthy, context-rich insight
that continuously retrains itself, explains its own uncertainty, and
feeds closed-loop control without human analysts.
Key Features: - Generative Telemetry Synthesis & Gap-Filling: Gen-AI
models learn multi-modal telemetry (packets, flow records, SNMP,
syslogs, DPI, spectrum scans) and hallucinate statistically faithful
“missing data” where sensors are sparse or silent, delivering 100 %
coverage at any time/space scale.
* Semantic Anomaly Narratives & Root-Cause Metrics: Instead of
threshold alerts, the model outputs human-readable stories
(“Between 02:13-02:19 UTC, TCP RTT on slice-C rose 38 % because 17
% of ECN-marked packets were re-routed via the Seattle POP due to
a mis-announced BGP community”). Each sentence is back-traced to
verifiable measurement samples.
* Self-Driving Measurement Campaigns: The AI translates high-level
intents (“tell me if user-perceived 4 K latency could exceed 150
ms during the next football final”) into dynamic sampler
schedules, probe paths, and packet structures; it launches the
campaign, stops when statistical confidence is reached, and
releases resources back to the data plane.
* Counterfactual & Predictive “What-if” Metrics: Given a proposed
config change (new AQM, additional slice, 400 GbE upgrade), the
generator produces the expected delay/loss/jitter distributions
before any byte is moved, letting operators compare KPI deltas
without real-world probing.
6. Use cases of Agentic AI for Operations & Management
Different use cases for Agentic AI on Operation & Management can be
identified, as described in the following sections.
6.1. Intent Based Service Delivery
Below is the diagram showcasing how network management AI agent takes
effect on the intent based service delivery process.
+----------------------------------------------------------+
| L3VPN Service Delivery Application |
+-------------------------+--------------------------------+
|
Intent |LPI
interface
|
+---------+ +------------V--------------------------------+
| | | |
|Knowledge| | Network Management AI Agent |
| Base <--> |
| | | +----------------------------------------+ |
+---------+ | | Intent Decomposing&Analysis | |
| | | |
+---------+ | | +----------++----------+ +------------+| |
| | | | | Config || Config | | Config || |
| Network | | | |Generation||Validation| |Distribution|| |
| Digital <--> | +----------++----------+ +------------+| |
| Twin | | +----------------------------------------+ |
| Tools | | |
+---------+ +---------------------------------------------+
+----------------------------------------------------------+
| Network Infrastructure |
+----------------------------------------------------------+
Legend: LPI - Language Programming Interface
Figure 2: Intent Based Service Delivery
Step a. L3VPN Service Delivery Application at the OSS layer uses
Language Programming Interface (LPI) to send service intent request
"Create L3VPN service with 2 VPN sites in London and Paris using L3SM
Service Model".
Step b. The Network Management AI Agent looks up knowledge base to
understand the intent and identify user's objective "VPN Service
Creation".
Step c. The Network Management AI Agent further interacts with
Knowledge base for expert experience and looks up thought of chain
related to "VPN Service Creation". And then the Knowledge base
returns results to the Network management AI Agent.
Step d. The Network Management AI Agent decomposes user intent and
break down the tasks into operational workflow including
configuration generation, configuration validation, configuration
distribution. For configuration validation, it will interact with
Network Digital Twin tools to obtain the validation results.
Step e. After L3VPN Service is delivered successfully, the Network
Management AI Agent will use LPI to return success results.
6.2. Cross-layer and Cross-domain Multi-Agent communication for
Complaint handling
In this scenario, automotive companies centrally collect complaints
from their customers (drivers) and use the operator’s complaint
system to feedback issues to the operator. The operator's BSS
trouble ticket system generates tickets from these complaints and
dispatches them to the OSS. The integrated vehicle networking
complaint handling agent within the OSS analyzes the trouble tickets
and performs fault localization. The ticket will be sent to the
corresponding vehicle networking trouble ticket agent within OSS
based on whether fault localization is within or beyond specific
maintenance domain.
The vehicle networking trouble ticket agent within the OSS will parse
the ticket into multiple multi-steps workflow and interact with the
IP network agent and mobile network agent within its management
domain to resolve the problem.
+-----------------------+
|Automobile Manufacturer|
| Complaints |
+-----------------------+
|
+----------------------+
| BSS Trouble tickets |
| System |
+----------------------+
|
+-------------------------------------------+
| OSS |
| |
| +---------------+ +----------------+|
| | Complaint | | Complaint ||
| | Handling Agent|-------| Handling Agent ||
| | In Domain A | | In Domain B ||
| +-------+-------+ +----------------+|
+---------+---------------------------------+
+--+-----------------------+
+----+---------+ +------+-------+
| Mobile | | IP |
| Network | | Network |
| Agent | | Agent |
+--------------+ +--------------+
Figure 3: IoV User Complaints Handling
o Tasks are triggered by natural language
* Complaints usually come from end-users or enterprises
- who may not have a deep understanding of network
o sometimes are unable to provide accurate descriptions
o Tasks possess both abstraction and expertise
* Abstraction: complaint content is unpredictable and the involved
domains cannot be anticipated
* Expertise: The final closed-loop of the task depends on the
network
o Tasks involve cross-layer and cross-domain aspects
* Cross-Layer: BSS/OSS -> Network
* Cross-domain:
- Technical domains (wireless network domain, backhaul network
domain)
- management & maintenance domains (i.e. across provinces and
cities)
6.3. AI Agent Driven Network Management
+-----------------+
| OSS |
+--------+--------+
|
|LPI
|
|
Model Invocation +---------+--------+
+--------------------------+ |
| Momery Access Network Management -----------+
| +-----------------+ AI Agent | |
| | +----| | |
| | | +-------^---+------+ |
| | | Response | |Execution |
| +----V---+ +----+-----+ +-+---V-----+ +------+------+
| | | | | | | | |
| | Memory +---|Tools-box <----| Validation+-----> Execution |
| | | | | | | | |
| +--------+ +----^-----+ +-----------+ +-------------+
| Model | Tool Calling Action Execution
| Invocation|
| |
+-V---------------------V------------------------------------------+
| Model Repository |
| (Task Decomposing, Reasoning, Data Analysis, Decision Making..) |
| +--------------+ +-----------------+ |
| | LLM Models | |Specialized Small| |
| +--------------+ | AI Models | |
| +-----------------+ |
+------------------------------------------------------------------+
Figure 4: AI Agent Driven Network Management
Traditional network operation and maintenance require extensive human
oversight and are constrained by predefined policies, limiting real-
time adaptability. Network management AI agents at the network level
enhance network intelligence and automation by integrating large
network foundation models, specialized small AI models, and feedback
closed loops mechanisms. The key functional requirements of the
Network management AI agent include:
* Integrate with large foundation models and specialized small
models for context-aware decision-making;
* Support Intent realizing including task decomposition,reasoning,
inference&prediction and decision making.
* Support autonomous execution of network service lifecycle
management, including network service delivery, network anomaly
detection, predictive maintenance and troubleshooting, network re-
optimization;
* Work with upper layer OSS to facilitate cross-layer collaboration,
enabling seamless communication between network elements;
7. Security Considerations
TODO Security
8. IANA Considerations
This document has no IANA actions.
9. References
9.1. Normative References
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119,
DOI 10.17487/RFC2119, March 1997,
<https://www.rfc-editor.org/rfc/rfc2119>.
[RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
May 2017, <https://www.rfc-editor.org/rfc/rfc8174>.
9.2. Informative References
[Agentic-AI-Wireless]
"Advanced Architectures Integrated with Agentic AI for
Next-Generation Wireless Networks", 2025,
<https://arxiv.org/html/2502.01089v3>.
[AIPREF] "IETF AIPREF WG", 2025,
<https://datatracker.ietf.org/group/aipref/about/>.
[ANIMA] "IETF ANIMA WG", 2025,
<https://datatracker.ietf.org/group/anima/about/>.
[RFC7575] Behringer, M., Pritikin, M., Bjarnason, S., Clemm, A.,
Carpenter, B., Jiang, S., and L. Ciavaglia, "Autonomic
Networking: Definitions and Design Goals", RFC 7575,
DOI 10.17487/RFC7575, June 2015,
<https://www.rfc-editor.org/rfc/rfc7575>.
Acknowledgments
TBA
Authors' Addresses
Yong-Geun Hong
Daejeon University
62 Daehak-ro, Dong-gu
Daejeon
34520
South Korea
Email: yonggeun.hong@gmail.com
Joo-Sang Youn
DONG-EUI University
176 Eomgwangno Busan_jin_gu
Busan
614-714
South Korea
Email: joosang.youn@gmail.com
Qin Wu
Huawei
101 Software Avenue, Yuhua District
Jiangsu
210012
China
Email: bill.wu@huawei.com
Benoit Claise
Everything OPS
Belgium
Email: benoit@everything-ops.net