Markdown Version | Transcript | Recording 1 | Recording 2 | Session Materials

Session Date/Time: 16 Mar 2026 08:30

NMRG

IETF 125 - NMRG Session A Chairs: Jean-François Peligry, Jefferson Campos Nobre

Summary

The Network Management Research Group (NMRG) met to discuss the intersection of Artificial Intelligence (AI), specifically Large Language Models (LLMs) and Agentic AI, with network management. Key topics included intelligent configuration management, automated protocol testing frameworks, the applicability of Model Context Protocol (MCP) and Agent-to-Agent (A2A) communication in networking, and updates to the Intent-Based Networking (IBN) use cases draft.

Key Discussion Points

1. Chairs Slides

• Jean-François opened the session, reminding participants of the IRTF’s research focus (long-term issues) vs. the IETF’s engineering focus.
• Announced an interim meeting on Saturday (joint with ETSI ZSM).

2. Towards Intelligent Network Configuration Management with LLMs​

• Wenlong Ding presented research on using LLMs for intent recovery, configuration generation, and verification.
• Technical Highlights:
  • Introduction of three engines: Semantic (IPAM mapping), Provenance (reverse tracing config), and Temporal (Git evolution/history analysis).
  • Proposed an agent-based verification mode that outperformed formal tools like Batfish in scalability by using LLMs to infer routing states.
• Discussion: Muhammad inquired about time savings for human operators. Wenlong Ding noted that for large scales where human analysis is infeasible, the system completes holistic analysis in minutes/hours, and is up to 20 times faster than state-of-the-art intent-based rules.

3. Agentic AI for Intent-Based Network Management

• Abdelkader Mekraouech introduced OSSGPT, a framework using multiple LLM agents (Assistant, Planner, Executor, Reporter) to translate natural language intents into API calls and JSON structures (specifically ETSI Network Service Descriptors - NSDs).
• Technical Highlights:
  • Fine-tuning of a "NSD Expert" model (Llama 3.2 3B) using LoRA.
  • Intent assurance loop using XGBoost for anomaly prediction and SHAP (Explainable AI) to identify root causes (e.g., CPU vs RAM exhaustion) for autonomous remediation.
• Discussion: A participant questioned the security and hallucination risks of autonomous agents. Abdelkader Mekraouech clarified that critical actions currently involve "human-in-the-loop" validation and Python-based semantic checking of JSON structures. Muhammad asked about the translation mechanism; Abdelkader explained that natural language is mapped to specific OSS API endpoints via the Planner agent.

4. Framework and Automation Levels for AI-Assisted Network Protocol Testing

• Yunze Wei updated the group on draft-wei-nmrg-ai-protocol-testing-01.
• Technical Highlights:
  • Shifted focus from "protocol understanding" to "protocol formalization"—a more rigorous process to translate RFCs into machine-readable formats.
  • Introduction of a hybrid test case generation method using templates and parameter decoupling.
  • Transitioned the code generation module to a multi-agent system.
• Discussion: Xing (CAICT) expressed concern about whether the testing methods are suitable for IETF standardization, suggesting a focus on the formalization and templates instead. Ken Chen asked about multi-protocol complexity; Yunze Wei acknowledged this as a future research area.

5. Applicability of MCP for the Network Management

• Natalie Roman presented draft-li-nmrg-mcp-applicability-network-management-02 regarding the Model Context Protocol (MCP).
• Technical Highlights:
  • Identified challenges in protocol design (stateful server-sent events) and security (prompt injection, tool poisoning).
  • Proposed four deployment scenarios: Standalone MCP servers, integrated controller-element communication, element-to-element, and third-party API consumption.
• Discussion: Minje asked about the relationship between MCP and Netconf/YANG. Natalie Roman explained that MCP servers can act as adapters to existing protocols. Haoran asked about network-specific threat models; Natalie indicated that general LLM threats currently apply, but network-specific research is needed.

6. Applicability of A2A to the Network Management

• Shailesh (on behalf of authors) discussed draft-cn-nmrg-a2a-applicability-network-management-02, focusing on Agent-to-Agent (A2A) communication.
• Technical Highlights:
  • Proposal to integrate IETF YANG-based structured data into A2A messages to provide machine-parsable semantics alongside natural language.
  • Integration of the "Network Incident" YANG model (from NMOP WG) into A2A messaging.
• Discussion: Xing asked about flexibility for non-YANG data. Shailesh noted that "agent cards" allow agents to expose specific capabilities, including whether they support YANG parsing or other structured formats.

7. Use Cases and Practices for Intent-Based Networking

• Chan-Ho Jung presented updates to draft-irtf-nmrg-ibn-use-cases-03.
• Technical Highlights:
  • Mapped existing use cases to the IBN methodology steps and the intent taxonomy defined in RFC 9315.
  • Added a new practice case: Cloud-based security systems using I2NSF (Interface to Network Security Functions).
  • Improved consistency in documentation across all nine use cases.

Decisions and Action Items

• The authors of draft-irtf-nmrg-ibn-use-cases have addressed chair comments and believe the document is nearing completion.
• Natalie Roman requested feedback on whether draft-li-nmrg-mcp-applicability-network-management is ready for Research Group adoption.

Next Steps

• IBN Use Cases: Chan-Ho Jung requested a Research Group Last Call (RGLC) following the next IETF meeting (Vienna).
• Protocol Testing: Authors of draft-wei-nmrg-ai-protocol-testing will focus on refining protocol formalization and explore potential collaboration with the BMWG.
• Interim Meeting: A joint session with ETSI ZSM is scheduled for Saturday following IETF 125.

Session Date/Time: 19 Mar 2026 08:30

NMRG

IETF 125 - Network Management Research Group (NMRG) Session II

Chairs: Jefferson Campos Nobre, Jerome Francois
Secretaries: Pedro Martinez-Julia, Cao-Wei Cee

Summary

The second NMRG session at IETF 125 focused on the intersection of Network Digital Twins (NDT) and Agentic AI. Key themes included the transition from traditional Large Language Models (LLMs) to reasoning and action-oriented agents, the offloading of inference to the data plane, and the necessity of human-in-the-loop architectures for safety and determinism. The chairs also announced a joint interim meeting with ETSI ZSM to further explore Agentic AI in network management.

Key Discussion Points

1. Intelligence Distribution Network (IDN) and Decentralized LLM Inference

Presenter: Hanling Wang Slides: Chairs slides (Includes content for IDN and Decentralized LLM)

• IDN Framework: Proposed shifting the network paradigm from content delivery (CDN) to intelligence distribution. The goal is to cache popular intelligence services at the edge to reduce latency and cloud dependence.
• Decentralized Inference: Presented a framework inspired by blockchain for scalable LLM inference. Key components include a layer-aware transport protocol and an economic protocol for heterogeneous participants.
• Discussion: Jefferson Campos Nobre questioned the fit of these drafts within NMRG compared to CATS (Computing-Aware Traffic Steering). Hanling Wang noted that while CATS focuses on routing, NMRG is better suited for the organization and management of these intelligent service frameworks.

2. Network Digital Twins (NDT) Hands-on Experience

Presenter: Marco Slides: Chairs slides (Placeholder for NDT presentation)

• Implementation: Shared results from testing a production-grade 5G core in an AWS/Kubernetes environment. Used Transformer and Prophet models to predict CPU/memory utilization and "cascade failures" where one microservice failure impacts others.
• Discussion: Chang (China Mobile) inquired about the latency of Transformer-based models for real-time operations. Marco clarified that while training is intensive, inference is fast enough for proactive reconfiguration in a closed-loop system.

3. Agentic AI for Intent-Based Network Management

Presenter: Chin Slides: Agentic AI for Intent-Based Network Management

• Evolution: Discussed the move toward Level 4 autonomous networks using "Large Reasoning Models" and "Large Action Models."
• Architecture: Proposed an "Agent Fabric" connecting network element agents, network agents, and service agents.
• Discussion: A participant (Speaker 1) cautioned against ignoring 20 years of research on autonomous agents and noted the need for better governance mechanisms to manage collectives of agents with potentially emerging conflicting goals.

4. Evaluating LLM Agents for Network Configuration

Presenter: Shuguang Slides: Towards Intelligent Network Configuration Management with LLMs​-Wenlong Ding

• Framework: Introduced a standardized evaluation benchmark using the Model Context Protocol (MCP) to interface between agents and network environments (GNS3).
• Metrics: Agents are scored on reasoning similarity, command accuracy (F1 score), and functional pass rates in the emulator.
• Discussion: Discussion centered on the need for a community-driven, diverse dataset. Laurent Ciavaglia suggested aligning with the GSMA Open Telco benchmark.

5. In-Network Inference Protocol (INIP)

Presenter: Chin (on behalf of Guorui Xie) Slides: Framework and Automation Levels for AI-Assisted Network Protocol Testing

• Concept: To reduce latency, neural networks are distilled into decision trees and deployed directly onto data plane devices (P4 switches) using match-action tables.
• Discussion: Jerome Francois noted similarities to prior "In-Network Computing" research and asked how this specifically differentiates itself. Chin clarified that the focus is on the distillation process for limited data plane resources.

6. Agentic IP Networks and Gateways

Presenter: Shuyi Slides: Applicability of A2A to the Network Management

• Agentic Gateway: Described a gateway that acts as an agent, performing local decision-making, protocol conversion, and sensing.
• Architecture: Outlined a 6-layer agentic network architecture including environment perception and A2A (Agent-to-Agent) protocols.

7. Architecture Principles for Autonomous Computer Networks

Presenter: Mohamed Slides: slides-125-nmrg-mcp-applicability-network-management-02

• Problem: Current AI network solutions are vendor-locked silos.
• Solution: Proposed a stack where each layer has a dedicated agentic network. Emphasized the need for standardized naming, discovery, and "guard-railing" to move from stochastic AI behavior to deterministic network outcomes.

8. Human-in-the-Loop and Security in Agentic Management

Presenter: Minze Slides: Use Cases and Practices for Intent-Based Networking

• Interface: Proposed a framework where agents provide explainability logs and confidence scores, allowing human operators to audit configurations before they are issued.
• Security: Highlighted threat vectors including prompt injection (modifying hostnames or deleting BGP sessions) and agent identity spoofing.

9. Agentic AI for Network Sensing

Presenter: Carlos

• Case Study: Explored using agents for distributed sensing. Identified open issues in agent discovery, registration, and conflict resolution in multi-domain sensing environments.

Decisions and Action Items

• Interim Meeting: A joint meeting with ETSI ZSM is scheduled for Saturday to discuss Agentic AI and its relationship with network management. Onsite participation only.

Next Steps

• Terminology: Several presenters and participants (including Mohamed) highlighted the urgent need for NMRG to agree on a common terminology for Agentic AI in network management.
• Collaboration: Authors of the evaluation framework and Agentic AI architectures are encouraged to share datasets and refine architectural layers on the mailing list and GitHub.
• Draft Progression: Presenters of the new frameworks (IDN, Agentic Gateway, and LLM evaluation) to incorporate feedback regarding prior autonomous networking research and refine their positioning within the research group.

Related Documents

draft-cn-nmrg-a2a-applicability-network-management-02, draft-irtf-nmrg-ibn-use-cases, draft-irtf-nmrg-ibn-use-cases-03, draft-li-nmrg-mcp-applicability-network-management, draft-li-nmrg-mcp-applicability-network-management-02, draft-wei-nmrg-ai-protocol-testing, draft-wei-nmrg-ai-protocol-testing-01