**Session Date/Time:** 24 Jul 2026 07:00

# [NMRG](../wg/nmrg.html)

## Summary
The Network Management Research Group (NMRG) met at IETF 126. The session focused heavily on the emerging intersection of artificial intelligence, large language models (LLMs), autonomous agents, network digital twins (NDTs), and automated intent-based network management. The group discussed several active internet-drafts and research presentations covering self-supervised closed loops, transformer-based anomaly detection, ontology reconciliation, agentic conflict resolution, intelligent data planes, prompt normalization schemas, and network knowledge graphs.

---

## Key Discussion Points

### 1. Introduction and Administration
* **Presenter**: Jérôme François (Co-Chair)
* **Slides**: [Intro slides](https://datatracker.ietf.org/meeting/126/materials/slides-126-nmrg-intro-slides-00)
* **Details**: Jérôme François presented the IETF Note Well, intellectual property rights (IPR) policies, and the code of conduct. He reminded participants that NMRG is an IRTF research group focusing on long-term research issues rather than near-term standards development, though it can publish informational or experimental RFCs.

### 2. Autosupervised Automation Loop System for Cloud-Native Infrastructures
* **Presenter**: Stefano Secci
* **Slides**: [Autosupervised Automation Loop System for Cloud-Native Infrastructures​](https://datatracker.ietf.org/meeting/126/materials/slides-126-nmrg-autosupervised-automation-loop-system-for-cloud-native-infrastructures-00)
* **Details**: Stefano Secci presented a MAPE-based closed-loop automation system developed by LIP6 and Nerotel.ai, targeted at 5G and cloud-native infrastructures (such as Open5GS and Open Air Interface). The system features an auto-supervised data pipeline that avoids manual labeling. An encoder-decoder (e.g., LSTM) is trained to minimize reconstruction errors on normal states. Anomaly detection outputs an anomaly fingerprint (an array of reconstruction errors across features or feature groups), which a Deep Reinforcement Learning (DRL) agent uses to navigate a multidimensional space and select mitigation/reconfiguration actions.
* **Discussion**:
  * John-Fritz asked whether user equipment (UE) data is utilized. Stefano responded that while personal UE data is not collected, they extract radio front-end KPIs on a per-UE level and are looking into vehicular/infrastructure UEs (e.g., trains or connected cars).
  * John-Fritz asked about integration with 3GPP's Network Data Analytics Function (NWDAF). Stefano clarified that while NWDAF primarily processes 3GPP core KPIs, this system goes beyond by scraping container-level metrics, physical server statistics, and radio front-end data.
  * In the chat, Jérôme François asked about the reinforcement learning observation space. Stefano clarified that the RL agent navigates a space with dimensions equal to the anomaly fingerprint array size, using reconstruction errors as inputs.

### 3. Building an Intelligent Network @SURF
* **Presenter**: Peter Boers
* **Slides**: [Building an Intelligent Network @SURF](https://datatracker.ietf.org/meeting/126/materials/slides-126-nmrg-building-an-intelligent-network-surf-00)
* **Details**: Peter Boers shared practical experiences from SURF (the Dutch National Research and Education Network) in deploying AIOps. The network transports 1.6 exabytes annually and is fully orchestrated. He discussed two components: *Sherlock* (anomaly detection/inferencing) and *Clippy* (autonomous troubleshooting agent). Sherlock evolved from using Long Short-Term Memory (LSTM) networks—which suffered from high retraining overhead due to seasonality—to utilizing an *Inverted Transformer* (iTransformer) architecture. The transformer model runs as an Oracle HTTP endpoint, can predict 32 steps ahead, handles attention between multiple correlated variables, and only requires training once or twice a year.
* **Discussion**:
  * Frank Brockners asked how the selected model compares to standard Time Series Foundational Models (TSFMs) like TimesFM, Moirai, Chronos, or Cisco's time series model. Peter responded that they are planning to evaluate TOTO. In the chat, Frank shared links to Cisco's [time-series-model](https://github.com/splunk/cisco-time-series-model) benchmarks and Salesforce's [GIFT-Eval](https://huggingface.co/spaces/Salesforce/GIFT-Eval).
  * Thomas Graf asked about handling seasonality (e.g., sports events) and suggested integrating the BGP Monitoring Protocol (BMP) to correlate traffic anomalies with routing table changes. Peter agreed that integrating routing-level events is a planned complexity.
  * Diego Lopez asked about the use of unlabeled datasets and the potential for digital twin emulation. Peter confirmed that the ultimate ambition is to expand the system into a network digital twin.
  * In the chat, Leandro Bertholdo asked whether Sherlock could explain anomalies (explainable AI) or if they had considered using an LLM to explain why an anomaly was detected.

### 4. Ontology Reconciliation
* **Presenter**: Christopher Janz
* **Slides**: [draft-janz-nmrg-ontology-reconciliation](https://datatracker.ietf.org/meeting/126/materials/slides-126-nmrg-draft-janz-nmrg-ontology-reconciliation-00)
* **Details**: Christopher Janz introduced [draft-janz-nmrg-ontology-reconciliation](https://datatracker.ietf.org/draft-janz-nmrg-ontology-reconciliation/), focusing on how cognitive agents can dynamically reconcile differing data models and ontologies on the fly without human intervention. The concept shifts from rigid, pre-standardized Yang structures to referenceable, flexible "lexicons" (nomenclature) augmented with glosses/explanations.
* **Discussion**:
  * Ken Watson asked for a practical example of the "table 1 vs. table 7" mapping. Christopher and Diego Lopez explained scenarios involving multi-provider boundary handoffs (e.g., a public carrier communicating with a private enterprise network) or internal silos (e.g., radio engineering vs. optical fiber teams) where data models do not align natively.
  * Alex Clemm asked via the mic whether this approach handles constraints and dependencies defined in Yang models. Christopher noted that it is early in the research and they are starting with basic structural translations before tackling complex logic.

### 5. Inter-Agent Conflict Resolution
* **Presenter**: Christopher Janz
* **Slides**: [draft-janz-nmrg-inter-agent-conflict-resolution](https://datatracker.ietf.org/meeting/126/materials/slides-126-nmrg-draft-janz-nmrg-inter-agent-conflict-resolution-00)
* **Details**: Christopher Janz presented [draft-janz-nmrg-inter-agent-conflict-resolution](https://datatracker.ietf.org/draft-janz-nmrg-inter-agent-conflict-resolution/), detailing the inevitability of conflicts when autonomous agents independently execute changes. The draft outlines 10 coordination families ranging from normative regulation and centralized arbitration to market-based allocations and Distributed Constraint Optimization (DCOP). A digital twin acts as a key pre-commit safety environment.
* **Discussion**:
  * Laurent Ciavaglia asked about remaining research gaps. Christopher replied that while cooperative approaches like DCOP are mathematically bound, applying market-based methods to intra-domain network management remains a highly active and challenging research area.

### 6. NaaS Agentic Negotiation
* **Presenter**: Christopher Janz
* **Slides**: [draft-janz-nmrg-naas-agentic-negotiation](https://datatracker.ietf.org/meeting/126/materials/slides-126-nmrg-draft-janz-nmrg-naas-agentic-negotiation-00)
* **Details**: Christopher Janz presented [draft-janz-nmrg-naas-agentic-negotiation](https://datatracker.ietf.org/draft-janz-nmrg-naas-agentic-negotiation/), showing how BSS/OSS boundaries can negotiate intents dynamically. He asserted that new, complex Yang models are not required to support dynamic, on-demand services; instead, existing models should be combined with negotiation process flows and a "digital wallet" containing enterprise spend boundaries.
* **Discussion**:
  * Jérôme François asked how NMRG can support this. Christopher suggested that since AI Ops and agentic networking are forming across side meetings and working groups, these topics should be brought to the new AI Ops mailing list to determine where the work should land.

### 7. Agentic AI for Intent-Based Networking
* **Presenter**: Alexander Clemm
* **Slides**: [draft-cxxx-nmrg-ai4ibn-00](https://datatracker.ietf.org/meeting/126/materials/slides-126-nmrg-draft-cxxx-nmrg-ai4ibn-00-00)
* **Details**: Alexander Clemm introduced [draft-cxxx-nmrg-ai4ibn](https://datatracker.ietf.org/draft-cxxx-nmrg-ai4ibn/), which revisits RFC 9315 concepts in the context of LLMs and agentic AI. The draft maps cognitive capabilities to the IBN lifecycle, highlighting translation, refinement, and orchestration planning as strong fits for AI, while pointing out severe constraints regarding high-velocity telemetry data processing.
* **Discussion**:
  * Toerless Eckert invited the community to contribute practical, real-world examples of AI elements within IBN systems to capture the current state of implementation in the document.
  * In the chat, Laurent Ciavaglia, Christopher Janz, and Jérôme François noted strong touchpoints between this draft and Christopher's ontology/lexicon work.

### 8. Intelligent Data Plane (IDP) Framework
* **Presenter**: Shang Yu
* **Slides**: [draft-xu-nmrg-idp-framework-00](https://datatracker.ietf.org/meeting/126/materials/slides-126-nmrg-slide-draft-xu-nmrg-idp-framework-00-00)
* **Details**: Shang Yu presented [draft-xu-nmrg-idp-framework](https://datatracker.ietf.org/draft-xu-nmrg-idp-framework/), which proposes an architectural framework to safely run AI-based decision functions on programmable data plane devices at line-rate. IDP ensures that processing is bounded (resource limits), policy-compliant, and observable. The framework defines logical Northbound, East-West, and Internal interfaces, along with an escalation process for low-confidence outputs.

### 9. Normalizing Multi-Vendor Network Inputs for LLM-Assisted Network Management
* **Presenter**: Shailesh Prabhu
* **Slides**: [draft-prabhu-nmrg-prompt-schema-llm-00](https://datatracker.ietf.org/meeting/126/materials/slides-126-nmrg-framework-for-normalizing-multi-vendor-network-inputs-for-llm-assisted-network-management-00)
* **Details**: Shailesh Prabhu introduced [draft-prabhu-nmrg-prompt-schema-llm](https://datatracker.ietf.org/draft-prabhu-nmrg-prompt-schema-llm/), addressing the challenge of feeding heterogeneous, multi-vendor syntax (e.g., CLI, Syslog) into LLMs. The draft proposes a normalization layer containing a Small Language Model (SLM) classifier to identify input types (performance, config, operational), which then formats the data into a standardized JSON, XML, or Yang schema before generating the LLM prompt.
* **Discussion**:
  * Laurent Ciavaglia and Roberta Robert expressed strong interest, noting that neutral schema normalization is a critical piece of the emerging "Agent Ops" and "AI Ops" domain.
  * Diego Lopez suggested expanding the draft with research on the most suitable tools for semantic mapping (e.g., knowledge graphs, terminological glossaries, cross-terms).

### 10. Network Digital Twin (NDT) Agent Architecture
* **Presenter**: Cheng Su
* **Slides**: [draft-wmz-nmrg-agent-ndt-arch-04](https://datatracker.ietf.org/meeting/126/materials/slides-126-nmrg-draft-irtf-nmrg-network-digital-twin-arch-01)
* **Details**: Cheng Su presented the fourth version of [draft-wmz-nmrg-agent-ndt-arch](https://datatracker.ietf.org/draft-wmz-nmrg-agent-ndt-arch/), which has transitioned from a multi-agent system to a hybrid multi-agent architecture. It integrates a single Northbound agent, human-in-the-loop safety checkpoints, knowledge graphs, and a Network Digital Twin (NDT) as a risk-free simulation sandbox. He presented a recent Hackathon demo where root cause analysis was validated using a Qwen-72B model.
* **Discussion**:
  * Cheng Su requested Research Group adoption of the draft. Jérôme François noted the request and urged community members to provide feedback on the mailing list.
  * In the chat, Christopher Janz raised concerns regarding potential overlap between this draft and the Network Management Architecture (NMA) draft in the NMOP working group, suggesting that they coordinate to ensure architectural complementarity.

### 11. Network Knowledge Graph Gateway
* **Presenter**: Sheng Cui
* **Slides**: [draft-nmop-cui-nkg-gateway-00](https://datatracker.ietf.org/meeting/126/materials/slides-126-nmop-cui-nkg-gateway-00-00) *(Note: This slide deck is referenced as draft-nmop-cui-nkg-gateway-00-00 in the session materials)*
* **Details**: Sheng Cui introduced [draft-nmop-cui-nkg-gateway](https://datatracker.ietf.org/draft-nmop-cui-nkg-gateway/), which tackles the usability barrier of Network Knowledge Graphs (NKGs). Operators typically lack expertise in graph-specific query languages (e.g., Cypher, SPARQL). The draft proposes a three-layer Gateway architecture (Upstream, Midstream, Downstream) to translate natural language intents into intermediate structural representations and execute low-level graph queries.

---

## Decisions and Action Items
* **Feedback on Drafts**: The Co-Chairs called for increased community feedback on the mailing list for all presented drafts, specifically:
  * [draft-wmz-nmrg-agent-ndt-arch](https://datatracker.ietf.org/draft-wmz-nmrg-agent-ndt-arch/) (following the presenter's request for Research Group adoption).
  * [draft-prabhu-nmrg-prompt-schema-llm](https://datatracker.ietf.org/draft-prabhu-nmrg-prompt-schema-llm/).
  * [draft-xu-nmrg-idp-framework](https://datatracker.ietf.org/draft-xu-nmrg-idp-framework/).

---

## Next Steps
* **AI Ops Mailing List Coordination**: Participants interested in Agentic AI, prompt schemas, and ontology reconciliation (including Christopher Janz, Shailesh Prabhu, Laurent Ciavaglia, and Alexander Clemm) will utilize the newly established **AI Ops mailing list** to coordinate cross-group activities and align agent architectures across NMRG, ANIMA, and NMOP.
* **IBN and Autonomic Networking Collaboration**: Alexander Clemm and Toerless Eckert will continue seeking practical use cases and industry implementations to incorporate into [draft-cxxx-nmrg-ai4ibn](https://datatracker.ietf.org/draft-cxxx-nmrg-ai4ibn/).