| Internet-Draft | Agent implications for YANG | July 2026 |
| Ma & Wu | Expires 7 January 2027 | [Page] |
The industry today is actively exploring the application of agentic AI to autonomous network operations. However, there is little attention given to the impacts that the introduction of agentic AI may impose on YANG modeling, and how YANG can be better leveraged to support AI-driven autonomous operations.¶
Based on some end-to-end intent translation workflow analysis that engage YANG models across service, network, and device layers, this document identifies some critical gaps when AI agents generate and operate on YANG data. The purpose of this document is not to substantially redesign YANG or define a new data model language. Instead, it aims to facilitate discussion on how existing YANG ecosystems can be better leveraged in the context of agentic AI, and how complementary mechanisms can bridge these gaps while preserving YANG’s interoperability.¶
This note is to be removed before publishing as an RFC.¶
The latest revision of this draft can be found at https://QiufangMa.github.io/AgentImpact2YANG/draft-mw-nmop-yang-ai-challenges.html. Status information for this document may be found at https://datatracker.ietf.org/doc/draft-mw-nmop-yang-ai-challenges/.¶
Source for this draft and an issue tracker can be found at https://github.com/QiufangMa/AgentImpact2YANG.¶
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 7 January 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.¶
RFC 8969 [RFC8969] establishes a framework for automating service and network management using YANG [RFC7950]. Combined with network management protocols such as NETCONF [RFC6241] and RESTCONF [RFC8040], YANG delivers deterministic and interoperable capabilities for configuration manipulation and state retrieval across service, network, and device layers.¶
Recently, Large Language Models (LLMs) and AI Agents have emerged as new enablers for network automation. Multiple IETF activities are ongoing in this area, proposals include but are not limited to:¶
The network digital twin [I-D.irtf-nmrg-network-digital-twin-arch] and agentic AI based architecture for AI driven network operations described in [I-D.wmz-nmrg-agent-ndt-arch], which deploys hybrid agent systems within the network autonomous domain, and allows agents to invoke network operational function modules and collaborate to solve non-trivial tasks.¶
The Network Management Agent (NMA) concept presented in [I-D.zhao-nmop-network-management-agent], which proposes that network management agents can be deployed at both orchestrator and controller layers, with A2A, A2C, and A2N interfaces enabling communication among agents to agents, controllers, and network elements.¶
The applicability of MCP [MCP] for Network Management [I-D.yang-nmrg-mcp-nm], which explores how the Model Context Protocol can refactor network management operations and network capabilities as LLM-callable tools.¶
The applicability of A2A to Network Management [I-D.yang-nmrg-a2a-nm], which outlines how the Agent to Agent protocol can be leveraged to develop various rich AI driven network applications, realize intent-based network management automation in the multi-vendor heterogeneous network environment.¶
While the industry is actively exploring the application of agentic AI to autonomous network, little attention has been paid to how YANG modeling, as the very foundation of model-driven network automation, would be impacted, and if there is any critical gaps that remain unrecognized/unsolved at the intersection of probabilistic AI agents and deterministic structured YANG.¶
This document analyzes two typical network operation workflows driven by hierarchical AI Agents: service provisioning/optimization and network troubleshooting. It identifies key gaps arising from the adoption of YANG data models in these scenarios. The purpose of this document is not to substantially redesign YANG or define a brand-new data modeling language. Rather, it seeks to adapt YANG for agentic AI. By identifying fundamental gaps, this document aims to facilitate community discussion on how to bridge the gap between them.¶
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.¶
The document uses the following terms defined in [RFC8309] and [RFC8969]:¶
The document uses the following terms defined in [I-D.hong-nmrg-agenticai-ps]:¶
Agentic AI¶
The following definitions are used throughout this document:¶
An autonomous software system driven by an artificial intelligence reasoning engine (typically a large language model) that integrates a set of resources such as data and tools to perceive its operational environment, dynamically plan and execute decisions, and invoke actions to achieve specific goals.¶
Unlike traditional deterministic or rule-based autonomous systems, an AI Agent exhibits non-deterministic reasoning, maintains runtime and historical state, and possesses the capability to dynamically decompose abstract high-level intents into a sequence of operational actions.¶
The property whereby an AI Agent may produce different outputs (e.g., generate different configuration or give different network diagnostic conclusions) across multiple invocations given the identical input, due to the probabilistic nature of the underlying model, historical context drift, or non-producible reasoning paths.¶
This section uses a simplified layered agentic AI reference architecture compliant with [I-D.wmz-nmrg-agent-ndt-arch] and describes two typical workflows: the network service provisioning/optimization and the fault troubleshooting. The workflows cover scenarios from configuration generation to YANG modelled data consumption, which are essential to expose and derive the gaps in Section 4. Note that these flows are not the only possible ones. The intent flows shown here are just examples, i.e., typical workflows that illustrate how YANG models at the service, network, and device layers can be engaged in an agent-integrated network automation context.¶
This document targets a hierarchical agent deployment architecture as specified in Figure 1. It is a simplified view that hides task agents, function modules or toolsets that are available for agents to invoke. The architecture may also use memory management (e.g., short-term and long-term memory) to maintain conversational continuity and enable contextual search. Refer to [I-D.wmz-nmrg-agent-ndt-arch] for the complete architecture design.¶
+--------------------------+
|Orchestrator |
|+------------------------+|
||Orchestrator AI Agent(s)||
|+-----------^------------+|
+------------+-------------+
|
|A2A Protocol
|//YANG-structured payload; or domain-level intent
|
+----------+----------+
|Controller| |
|+---------v---------+|
||Network AI Agent(s)||
|+-------------------+|
+----------^----------+
|
|NETCONF/RESTCONF/Telemetry
|
+----------v----------+
| Network Devices |
+---------------------+
A complete end-to-end workflow focusing on AI agents generating configuration from high-level intents is defined as follows.¶
A network operator submits an end-to-end service request expressed in natural language including explicit service requirements. The intent may imply some constraints such as not affecting existing network services.¶
The orchestrator AI agent parses the natural language intent, maps it into the corresponding service model and parameters for user confirmation, and then decomposes it into structured configuration using network model.¶
YANG-structured data are encapsulated as message payload, and transmitted to network AI agents along with other context metadata via agent communication protocol like Agent-to-Agent (A2A).¶
The network AI agent converts received tasks into structured device model layer operations. The configuration is handed off to the network controller's Southbound Interface (SBI) and delivered via YANG-driven protocols such as NETCONF and RESTCONF. Validation checks may be performed before the configuration is delivered to network devices.¶
Network devices receive incoming configuration, execute changes, and report execution response.¶
A network operator submits a high-level operational intent, requesting the network AI agent to continuously monitor the network, perceive emerging anomalies, perform active diagnosis, and conduct autonomous repair while reporting root causes and handling results.¶
Network AI agent grasps a holistic view of the network operational state across multiple network devices and YANG modules. Instead of flooding the Network AI agent with massive, raw streaming telemetry data which would instantly overwhelm the LLM's context window and incur prohibitive token costs, agent may leverage existing network data analytics components or tools to be aware of real-time network state. Network AI agent could detect network anomalies swiftly, which enables the prompt identification of potential issues before they escalate into major faults.¶
Once an issue or fault is identified, network AI agent diagnoses the exact cause and generate targeted repair solutions. These solutions may involve making configuration changes on relevant network devices, for example, when a failure is caused by misconfiguration.¶
The network AI agent autonomously executes the formulated repair actions by delivering corresponding YANG configuration changes to devices, after invoking the network digital twin to simulate the proposed repair solution successfully.¶
In real-world network operations, operator intents often contain expectations and constraints, relaxation preferences, retry strategies, or fallback rules. Nevertheless, YANG is designed solely focuses on defining data structures and formats, syntax and semantic rules and static data dependencies. It lacks the ability to express dynamic and flexible operational logic, which creates a hard boundary for carrying intent and limits the decision space of multi-agent system.¶
This Gap hinders step 3 of the service provisioning/optimization workflow (Section 3.2). To adapt to static YANG structures, orchestrator AI agents have to prune and simplify the original intent. Implicit operational requirements embedded in original intents, such as service continuity, high availability and security compliance, are likely lost during the transformation process.¶
This premature semantic pruning deprives downstream network AI agents of the flexibility to make dynamic decisions, perform graceful degradation or apply fault-tolerant adjustments when network runtime status changes.¶
Existing YANG serialization formats (e.g., XML and JSON [RFC7951]) are heavily optimized for either human readability or strict software parsing. When adapted to agentic AI environments, these formats introduce significant redundancy, and thus consumes substantial LLM context window capacity and incurring significant token costs. This gap is explicitly exposed during step 3 of the workflow in Section 3.2 and step 2 of the workflow in Section 3.3, where verbose encoding adds latency and cost without contributing semantic value.¶
Highly compressed binary alternatives like CBOR mitigate bandwidth issues but introduce a fundamental incompatibility with LLM processing pipelines: LLMs operate exclusively on text token streams and cannot consume raw binary payloads directly. Currently, there is a clear gap for a token-efficient, semantic-preserving encoding format specifically designed for AI-agent consumption.¶
when tasked with ambiguous instructions, missing/unreliable data, deep reasoning, or executing multi-step decisions, the agent's behavior could inevitably introduce randomness and unpredictability. These unpredictable variations in Agent and LLM outputs can significantly impact the reliability and consistency of network operations.¶
For the same high-level intent, an agent may generate different YANG instance data due to historical context drift or the probabilistic nature of the underlying large language models (steps 2 and 4 of the service provisioning/optimization workflow in Section 3.2). Likewise, when analyzing network anomalies or incidents, multiple distinct diagnostic trajectories and explanations may emerge even for the exact same observed network operational state (steps 3 and 4 of the network troubleshooting workflow in Section 3.3).¶
This section explores several operational directions to try to bridge the gaps identified in Section 4 and improve the AI-readiness of existing YANG ecosystem.¶
A promising direction for partially addressing gap in Section 4.3 is to refactor YANG‑based network operations into AI‑invocable tools using the Model Context Protocol (MCP). Instead of requiring an AI agent to generate raw YANG instance data directly, which is prone to hallucination or exhibits non-determined behavior, the agent invokes an MCP tool with structured parameters. By shielding agents from low-level syntax and model complexity, this abstraction layer reduces the agent’s burden of understanding complex YANG schemas and mitigates the risk of generating invalid configuration data.¶
The applicability of MCP to network management is discussed in [I-D.yang-nmrg-mcp-nm]. Industry open-source exploration is emerging towards this direction, e.g., gNMIBuddy [gNMIBuddy] provides a toolkit to wrap gNMI and OpenConfig YANG data models based network operations, designed primarily for LLMs with MCP integration.¶
Gaps identified in Section 4.1 can be partially addressed by adding a semantic layer on top of YANG data models. Knowledge graphs (KGs) provide a machine‑readable representation of network knowledge, enabling AI agents to better understand the semantics, relationships, and constraints embedded in the network.¶
IETF work in this area includes [I-D.mackey-nmop-kg-for-netops] and [I-D.tailhardat-nmop-incident-management-noria], which correlate data from different network planes, e.g., management, control, and data planes and present a holistic view of network status.¶
When a user expresses a high-level intent such as "check why the VPN tunnel is down", a KG‑enhanced agent can query the YANG-based knowledge graph to understand the relationships between relevant services and metrics. Furthermore, KGs can explicitly model concepts such as "preferred vs. optional", temporal KGs can model concepts such as "transient vs. persistent failure", which are currently absent from YANG, partially addressing Section 4.1.¶
There is also other work in NMOP focusing on semantic enrichment that could serve as a foundation layer for agentic AI driven network anomaly detection. A set of ongoing drafts include:¶
[I-D.ietf-nmop-network-anomaly-architecture] provides a reference architecture for knowledge based service disruption detection;¶
[I-D.ietf-nmop-network-anomaly-lifecycle] defines an experiment for managing the lifecycle process of a network anomaly detection system, spanning across detection, validation, and refinement.¶
[I-D.ietf-nmop-network-anomaly-semantics] describes common network symptom semantics across different network planes.¶
An AI agent gains a structured and machine-understandable vocabulary for describing, querying, and reasoning about network anomaly. The standardized anomaly semantic metadata provides the missing semantic hooks, and contributes to bridging the gap in Section 4.1.¶
To bridge the semantic loss challenge specified in Section 4.1 without breaking the backward compatibility with existing network infrastructure, this section outlines a complementary approach that has emerged from operational deployments, that is the use of a declarative intent and policy overlay.¶
This approach does not modify YANG or propose new YANG statements. Instead, it introduces a logical layer above YANG that serves as the bridge between human-readable intent and machine-executable configuration.¶
Take a natural language based service provisioning intent for example, Instead of forcing AI agents to directly translate loose, high-level operator intent into rigid, low-level structural service YANG leaf nodes, the agent could first compile the intent into a declarative intent/policy overlay. This overlay explicitly encapsulates invariants, business constraints, and dynamic behavioral policies that existing service models cannot naturally represent. The overlay is co-delivered alongside the generated concrete YANG instances to the lower-layer agents.¶
As noted in Section 4.2, the verbosity of YANG instance data, particularly in JSON/XML encodings, poses a practical barrier to cost-effective agent operations. One possible way forward is the development of a token-efficient YANG serialization mechanism tailored to agentic consumption.¶
The integration of agentic AI with YANG-based network automation introduces new attack surfaces and operational risks that differ from traditional deterministic network management.¶
For example, AI-generated configuration risks are introduced by agent autonomous decision-making. Unlike manually crafted or scripted configurations, LLM-generated configuration edits may contain unintended misconfiguration or semantic deviations. Such errors can lead to service outages, policy violations, or enlarged attack surfaces. The human-in-the-loop mechanism and network digital twin simulation and validation could mitigate this concern by blocking high-risk unvalidated changes before configuration is committed. Furthermore, tool invocation and credential exposure risks arise when AI agents are granted access to network devices via MCP servers. Open-source MCP implementations enable LLM clients to execute read/write operations over managed devices. Unauthorized or malicious prompt injection may trigger abusive command execution, configuration tampering, or information leakage. Token-based authentication, transport encryption, and tool-level guardrail mechanisms (blocked command lists and restricted configuration scopes) are required to enforce access control.¶
This document has no IANA actions.¶