| Internet-Draft | Agentic AI Framework | November 2026 |
| Jeong, et al. | Expires 20 May 2027 | [Page] |
This document specifies an agentic AI framework for the management of networks, security, and AI services. This framework is an Intent-Based System (IBS) to enforce a human-language intent at a target network with the translated policy for networks, security, or AI services. It consists of a User System, a Cloud System, an Edge System, and a Device. Target devices in this framework include Network Forwarding Elements (e.g., router and switch), Network Security Functions (e.g., firewall, web filter, and DDoS-attack mitigator), Internet-of-Things (IoT) Devices, Physical AI Robots, Software-Defined Vehicles (SDV), Unmanned Aerial Vehicles (UAV), and Urban Air Mobility (UAM). This document describes the agentic AI framework with its components and interfaces where the components include AI agents and supporting knowledge repositories (i.e., knowledge base), and also inter-agent interfaces use the Agent-to-Agent (A2A) protocol.¶
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 20 May 2027.¶
Copyright (c) 2026 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. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Agentic AI opens a new door providing human users with human-friendly services through Large Language Models (LLMs) such as GPT-3 [GPT-3] and Flan-T5 [Flan-T5]. Also, network softwarization has been widely adopted in various domains such as data centers and mobile networks [TS-23.501]. The network softwarization is realized with Network Functions Virtualization (NFV) [ETSI-NFV][ETSI-NFV-Release-2] and Software-Defined Networking (SDN) [RFC7149]. Furthermore, Intent-Based Networking (IBN) [RFC9315][Survey-IBN-CST-2023][ClickINC] can support human-friendly configuration for target networks and security systems.¶
This document specifies an agentic AI framework for the management of networks, security, and AI services. This framework is an Intent-Based System (IBS) including the IBN that enforces a human-language intent at a target network. It aims at the IBS in data centers, corporate networks, campus networks, and home networks rather than public networks like the Internet.¶
The agentic AI framework in this document supports the translation of a human-language intent and the enforcement of the corresponding service(s) through multiple AI agents in target networks. This intent can be translated into a specific policy for networks, security, or AI services. The agentic AI framework in this document consists of a User System, a Cloud System, an Edge System, and a Device. Devices under the control of this framework include Network Forwarding Elements (e.g., router and switch), Network Security Functions (e.g., firewall, web filter, and DDoS-attack mitigator), Internet-of-Things (IoT) Devices, Physical AI Robots, Software-Defined Vehicles (SDV) [AUTOSAR-SDV][Eclipse-SDV][COVESA], Unmanned Aerial Vehicles (UAV), and Urban Air Mobility (UAM). Thus, this document describes the agentic AI framework with its components and interfaces where the components are AI agents and the interfaces use the Agent-to-Agent (A2A) protocol [A2A][A2A-Specification].¶
This document uses the terminology described in [RFC9315] and [I-D.irtf-nmrg-ibn-usecases]. The new terms are defined below:¶
This section specifies an Agentic AI System for AI Service Management in target domains, as shown in Figure 1. In this figure, a human user (e.g., network administrator, building manager, and home user) uses a User System (e.g., smartphone and computer) for managing networks, security, and AI services. The User System interacts with a Cloud System or an Edge System for the human user's intents. A Cloud System manages multiple Edge Systems for the sake of human users. An Edge System manages multiple Devices for networks, security, and AI services like in Figure 1. The devices can be either Virtual AI Objects (VAOs) or Physical AI Objects (PAOs) having at least one Worker AI Agent for the interaction with an Edge Agent and other Worker AI Agents.¶
User System
+------------+
| User Agent |
+------------+
^ ^
| A2A | A2A
| v
Edge System v Cloud System
****************************************** **********
* * * *
* * * +----------+ *
* +-----------------+ * * | Cloud | *
* | Edge Agent | *<->* | Agent | *
* +--------^--------+ *A2A* +----------+ *
* | * * *
* v * * *
****************************************** **********
^ ^ ^
| A2A | A2A | A2A
v v v
+----------------+ +----------------+ +----------------+
| +------+ | | +------+ | | +------+ |
| |Worker| | | |Worker| | | |Worker| |
| |Agent1| |<-->| |Agent2| |<-->| |Agent3| |
| +---^--+ |A2A | +---^--+ |A2A | +---^--+ |
| | MCP | | | MCP | | | MCP |
| +--v----+ | | +--v----+ | | +--v----+ |
| |Modules| | | |Modules| | | |Modules| |
| +-------+ | | +-------+ | | +-------+ |
+----------------+ +----------------+ +----------------+
Device1 Device2 Device3
<----> Wired/Wireless Link
Figure 2 shows an Agentic AI Framework for the management of networks, security, and AI services. A User Agent in the User System generates an intent for the human user's service demand and sends it to either a Cloud System or an Edge System. Either a Cloud Agent in the Cloud System or an Edge Agent in the Edge System receives the intent from the User Agent.¶
In the case where the Cloud Agent receives an intent from the User Agent, the Cloud Agent translates the intent into the corresponding high-level policy through its Intent Translator. For the authentication and authorization of a User Agent, the User Agent in the User System needs to contact the Cloud Agent in the Cloud System. This security process is necessary for the management of networks, security, and AI services at the target domains where the User Agent has access permission. This is different from the assumption that an AI agent can discover other AI agents to help its task like travel agents to reserve airplane tickets, hotel rooms, and local transportation tickets.¶
<User System>
+----------------------+
| User Agent |
|(Smartphone, Computer)|
+----------------------+
^ | ^
| | Intent | Report
| v | <Cloud System>
| +----------------------------------------------------------------+
| | +-------------------+ Agent Information +--------------------+ |
| | | Cloud Agent |<----------------->| Knowledge Base | |
| | |(Intent Translator,| |(Edge Agent Catalog)| |
| | | Orchestrator) | | | |
| | +-------------------+ +--------------------+ |
| +----------------------------------------------------------------+
| ^ ^
| Intent | Intent (Report, Monitoring Data) | Agent Card
|(Report) | |
v v <Edge System> v
+------------------------------------------------------------------+
| +-------------------+ +----------------------+ |
| | Edge Agent | Agent Information | Knowledge Base | |
| |(Intent Translator,|<----------------->|(Worker Agent Catalog)| |
| | Orchestrator) | | | |
| +-------------------+ +----------------------+ |
+------------------------------------------------------------------+
^ ^
| Intent (Report, Monitoring Data) | Agent Card
| |
v <Devices with Worker Agent> v
+---------------------------------------------------------------------+
| <Device1> <Device2> <Device3> |
| +------------------+ +----------------------+ +-----------------+ |
| | Worker Agent1: | | Worker Agent2: | | Worker Agent3: | |
| | SDN Controller | | Security Controller | | Physical AI | |
| | (Router, Switch) | |(Firewall, Web Filter)| |(Robot, SDV, IoT)| |
| +------------------+ +----------------------+ +-----------------+ |
+---------------------------------------------------------------------+
Each Cloud Agent or Edge Agent accesses its Knowledge Base through a repository access interface. Its Intent Translator retrieves domain-specific schemas and semantic constraints from the policy-knowledge part of the Knowledge Base to support policy generation and validation. There exists an Agent Catalog that is a collection of Agent Cards [A2A-Specification] used for searching for candidate agents to perform the intent. There are two kinds of Agent Catalogs, namely an Edge Agent Catalog and a Worker Agent Catalog. The Edge Agent Catalog contains the Agent Cards of Edge Agents inside Edge Systems. On the other hand, the Worder Agent Catalog contains the Agent Cards of Workder Agents inside Devices. The Knowledge Base supplies information for intent translation and enforcement such as intent translation, policy validation, and agent selection.¶
Note that the agent discovery for an Agent Catalog can be performed in the following two approaches. The first approach is a reactive way using an agent discovery protocol (e.g., DNS and mDNS) [I-D.seethiraju-dawn-dan][I-D.zhang-dawn-agent-discovery-framework]. The second approach is a proactive way (e.g., a capability registration protocol) [I-D.ietf-i2nsf-capability-data-model][I-D.ietf-i2nsf-registration-interface-dm][I-D.ietf-i2nsf-applicability]. The Agentic AI Framework in this document can accommodate these two kinds of agent discovery approaches.¶
A policy-knowledge role is used to translate an intent into a specific intent (i.e., a refined intent) with the Policy Knowledge Base in [I-D.gu-nmrg-intent-translator]. Policy schemas and semantic constraints are supplied for each supported domain. For policy generation, a schema (i.e., a data model) with constraints for the service management needs to be developed. Note that a policy based on such a scheme can be expressed in XML [W3C.REC-xml-20081126], JSON [RFC8259], or YAML [YAML].¶
Once the User Agent is authenticated and authorized through the Cloud Agent, the User Agent can directly interact with an Edge Agent in an Edge System that can handle the User Agent's intent with its Work Agents, as shown in Figure 2.¶
When the Edge Agent receives a specific intent (i.e., a refined intent) from the Cloud Agent, it refines that intent into another specific intent or a high-level policy for Worker Agents. When it receives a natural-language intent directly from the User Agent, its Intent Translator first generates a structured high-level policy, which is then refined in the same way. Each specific intent represents a worker-specific goal or policy in an XML, JSON, or YAML document according to its target model. Within this framework, the Intent Translator queries the Agent Catalog to obtain candidate Agent Cards and evaluates their suitability for the requested task. Policy refinement and Worker Agent selection are extensions introduced by this document. Refinement preserves the goals and constraints of the original intent.¶
The Edge Agent sends the specific intent or the high-level policy to the selected Worker Agent(s) to enforce the requested service corresponding to the intent from the User Agent.¶
When a Worker Agent receives a specific intent or a high-level policy, it performs the requested service using its internal resources and external resources via MCP [MCP-Specification] with predefined APIs.¶
Operation results for the given intent are delivered as a report (e.g., success or failure) or monitoring data (e.g., alarms, events, logs, or counters) [I-D.ietf-i2nsf-nsf-monitoring-data-model] from a supervised source component (e.g., Worker Agent, Edge Agent, and Cloud Agent) to a supervising destination component (e.g., Edge Agent, Cloud Agent, and User Agent) in the Agentic AI Framework in Figure 2. When a Cloud Agent receives a report or monitoring data from an Edge Agent, the Orchestrator in the Cloud Agent checks the report or monitoring data to determine an appropriate action, depending on the message in the report or the monitoring data. Also, when an Edge Agent receives a report or monitoring data from a Worker Agent, the Orchestrator in the Edge Agent handles the report or monitoring data in the same way as the Cloud Agent. Therefore, the operation results (i.e., reports and monitoring data) can be used as feedback for fulfilling and assuring the intent through the closed-loop control in the Agentic AI Framework [I-D.irtf-nmrg-ibn-usecases].¶
Interfaces in the Agentic AI Framework use A2A [A2A-Specification] and MCP [MCP-Specification]. The interfaces among the User Agent, the Cloud Agent, the Edge Agent, and the Worker Agents use A2A, as shown in Figure 1. The message formats used by A2A conform to the specification of A2A [A2A-Specification]. The interfaces among agents and modules in each system (e.g., User Agent, Cloud Agent, Edge Agent, and Worker Agent) use MCP, as shown in Figure 1. The message formats used by MCP conform to the specification of MCP [MCP-Specification].¶
This document does not require any IANA actions.¶
Authentication and authorization mechanisms should be used to check the validity of User Agents. For trustworthy cooperation, mutual authentication should be performed beforehand among the Cloud Agent, the Edge Agents, and Worker Agents.¶
An intent from a User Agent should be evaluated by either the Cloud Agent or the Edge Agent to determine whether it is harmless to the Agentic AI Framework or not. If the intent is harmful, either the Cloud Agent or the Edge Agent should log this intent and report the harmfulness of the intent to the User Agent which issued the intent.¶
For reliable and agile AI services, the Agentic AI Framework should perform governance by monitoring and auditing the activities of AI agents including the User Agent, Cloud Agent, Edge Agents, and Worker Agents. Such an auditing system may operate as an independent entity in the Agentic AI Framework so that it can audit the AI agents without any bias.¶
The following changes are made from draft-jeong-nmrg-agentic-ai-framework-00:¶
This document is a product of the NMRG including the RG Chairs (Jerome Francois and Jeferson Campos Nobre). This document benefitted from the review and comments from the following people: Yong-Geun Hong and Joo-Sang Youn. The authors sincerely appreciate their efforts and kind help.¶
This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea Ministry of Science and ICT (MSIT) (No. RS-2024-00398199).¶
The following are coauthors of this document:¶