Paper deep dive

Agentic AI for IPoDWDM Network Lifecycle Automation: An MCP-Enabled Architecture

Chunmin Xia, Jakub Harbaczewski, Nikhil Dsilva, Julie Raulin, Dominic Schneider, Achim Autenrieth

Year: 2026Venue: arXiv preprintArea: cs.NIType: PreprintEmbeddings: 12

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%

Last extracted: 7/8/2026, 4:13:21 AM

Summary

This paper introduces a distributed, vendor-agnostic multi-MCP architecture for SDN-based automation and autonomous control of multi-layer IPoDWDM networks. The framework leverages an LLM-powered AI agent, specialized MCP servers, and optical telemetry to enable end-to-end service lifecycle management, closed-loop cross-layer control, and performance estimation, with experimental validation on a real testbed.

Entities (11)

Agentic AI Architecture · framework · 98%IPoDWDM Network · network · 97%Model Context Protocol (MCP) · protocol · 95%Adtran Networks SE · organization · 93%Mosaic Network Controller · sdncontroller · 92%Coherent Pluggables · hardware · 91%GNPy Model · software · 90%SUSTAINET-Advance · project · 90%NETCONF · protocol · 89%GPT5.2 · llm · 88%Redis · software · 85%

Relation Signals (8)

Agentic AI Architecture → automates → IPoDWDM Network

confidence 97% · enables fully automated, vendor-agnostic, multi-layer lifecycle management of IPoDWDM networks, spanning physical-layer setup through IP-layer provisioning

Agentic AI Architecture → uses → Model Context Protocol (MCP)

confidence 95% · model context protocol (MCP) enables scalable and composable development and integration of different kinds of tools in different specific domains

Mosaic Network Controller → developedby → Adtran Networks SE

confidence 93% · The Mosaic Network Controller (MNC), an SDN controller from Adtran

Agentic AI Architecture → employs → Mosaic Network Controller

confidence 92% · The Mosaic Network Controller (MNC), an SDN controller from Adtran, is used as the SDN control platform

IPoDWDM Network → consistsof → Coherent Pluggables

confidence 91% · interconnecting coherent pluggables from routers of different vendors including Juniper, Cisco and Arista

Agentic AI Architecture → fundedby → SUSTAINET-Advance

confidence 90% · This work has received funding from the German Federal Ministry of Research, Technology, and Space (BMFTR) project SUSTAINET-Advance

Agentic AI Architecture → utilizes → GNPy Model

confidence 90% · GNPy server is used to estimate and predict the E2E performance... open source GNPy model

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:We present a distributed, vendor-agnostic multi-MCP architecture for SDN-based automation and autonomous control of multi-vendor, multi-layer IPoDWDM networks. The framework enables E2E service lifecycle automation, closed-loop cross-layer control using GNPy model and optical telemetry, and is experimentally validated on a IPoDWDM testbed.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

11,193 characters extracted from source content.

Expand or collapse full text

Agentic AI for IPoDWDM Network Lifecycle Automation: An MCP-Enabled Architecture Chunmin Xia, Jakub Harbaczewski, Nikhil Dsilva, Julie Raulin, Dominic Schneider, Achim Autenrieth Adtran Networks SE, chunmin.xia@adtran.com Abstract We present a distributed, vendor-agnostic multi-MCP architecture for SDN-based automation and autonomous control of multi-vendor, multi-layer IPoDWDM networks. The framework enables E2E service lifecycle automation, closed-loop cross-layer control using GNPy model and optical telemetry, and is experimentally validated on a IPoDWDM testbed. ©2026 The Authors Introduction The rapid development and growing applications of artificial intelligence (AI), together with the ex- plosive growth of different data hungry applica- tions, is accelerating requirements for highly dy- namic, deep automation and increasing auton- omy imposed on IP-over-DWDM (IPoDWDM) networks. As an universal, standardized interface between autonomous AI agents and the real- world systems, model context protocol (MCP) [1,2] enables scalable and composable develop- ment and integration of different kinds of tools in different specific domains. Agentic AI application concepts focusing on some specific areas are seen in [3-9]. Based on both simulator/digital twin and real-world test beds, we propose and exper- imentally demonstrate an MCP-server-based agentic-AI architecture that enables fully auto- mated, vendor-agnostic, multi-layer lifecycle management of IPoDWDM networks, spanning physical-layer setup through IP-layer provision- ing and performance telemetry monitoring. Network setup and assumptions As shown schematically in Fig. 1, we consider a representative ring network with at least three network nodes, each comprising ROADMs, vari- able gain booster and pre-amplifiers as well as couplers/filters, interconnecting coherent plugga- bles from routers of different vendors including Juniper, Cisco and Arista as well as transponders from Adtran. It is necessary to point out our MCP servers can be easily extended to support any other routers/transponders with given necessary NETCONF configuration files. The Mosaic Net- work Controller (MNC), an SDN controller from Adtran, is used as the SDN control platform and is compatible with the available hardware testbed. AI agent and MCP tools development We have developed our own AI agent based on a Large Language Model (LLM) GPT5.2. The cli- ent prompts are saved in a memory datastore us- ing Redis and attached to the Agent together with any new requests to keep the memory of the con- text within one session. By evaluating streamable HTTP-based MCP servers and our own agent Fig. 1: Agentic AI assisted IPoDWDM network based on multiple MCP servers against commercial agents (e.g., Copilot, Claude), we identified limitations and iteratively improved our MCP servers and agent. In considering the importance of security and trustworthiness along with the AI applications, we keep always the harness engineering discipline in mind during developing all the MCP servers. First, all authentication credentials are managed through environment-based configuration param- eters for general use cases, while policy-driven Eclipse Dataspace Components Connector (EDCC) [10,11] mechanisms are used to meet higher-level security requirements. Secondly, in order to improve the transparency and ensure a positive user experience, each MCP tool is di- vided into as many atomic steps as possible, each of which is checked or validated carefully in order to return a clear meaningful information for any conditions even when an unexpected failure occurs. The MCP servers are developed and tested first of all based on digital twin like Adtran network simulators and then validated in real networks. Multi-MCP server architecture and workflow We aim to realize a deep smart automation of IP- oDWDM network with human language to AI Agent through creating different MCP servers working jointly in multiple network operation lay- ers from physical equipment setup, fiber connec- tions, device configuration to service provisioning and performance estimation as well as telemetry. As shown in Fig. 2, we have implemented and tested the following MCP servers: 1. Operation automation server: by a human prompt with a given SDN controller IP, any sub- network topology can be loaded automatically into a webpage. Nodes, shelves, cards and ports are all visualized for any further operations in- cluding fiber or nodes connections, device con- figurations and service creation, either through the GUI or the prompt or the combination of both. As indicated by the top part in Fig. 2, the user can prepare network operational data (e.g., fiber connections) via the GUI, visually review it, save it into batch files, and submit them to the Agent for processing with one prompt. All the creation- or deletion-requests can be verified by the corre- sponding show tool. 2. Coherent pluggables control server: Co- herent pluggables in routers or transponders are controlled directly by plug control MCP server, which realizes the following three fundamental functions: configuration, configuration status check or verification and performance metric monitoring after provisioning end-to-end (E2E) service. The detailed workflow is shown Fig. 3. Fig. 2: Workflow chart of MCP servers enabled IPoDWDM network Fig. 3: Workflow chart of coherent plug control MCP server An important observation is that the current state of the art of routers and pluggable optics from different vendors remains heterogeneous with respect to their support for NETCONF, OpenConfig, and gNMI protocols. Even with a same router vendor, the concrete NETCONF /openconfig XML RPC templates may be variable for different router models or different operating system (OS) software versions. Consequently, vendor information, hardware models, and OS versions were obtained via NETCONF or CLI and mapped to or inserted into vendor-specific Jinja2-based NETCONF XML templates (Fig. 3), enabling vendor-agnostic management by hiding vendor-specific differences from the agent. 3. Service provisioning server: Three differ- ent services provisioning MCP tools have been implemented: (1) based on Adtran cloud-based reconfigurable optical management & automation (croma) platform by creating the service connec- tions step by step from one ROADM degree to another. This is considered as a generic ap- proach and does not need SDN controller sup- port. (2) Create an alien service first between the ports from the OLS side that are connected to the router ports. (3) Single-step E2E service provi- sioning is currently supported on Juniper routers, with ongoing implementation to extend compati- bility to Cisco and Arista devices. 4. GNPy server: after service provisioning, GNPy server is used to estimate and predict the E2E performance. As shown in Fig. 2, in order to use the open source GNPy model [12], all the network elements together with their configura- tion and spec parameters in the service path from the real network topology and equipment configurations are firstly discovered and then converted into GNPy model compatible network .json format. Then the power and OSNR evolu- tion along the path can be calculated. 5. Telemetry server: All the performance met- rics of coherent transceivers can be retrieved based on queried data set and history time slot. Validation and demonstration Based on a real testbed shown in Fig. 4(a) and (b) by user prompts, two 400G DP-16QAM coher- ent optical pluggables (operated in 400ZR DWDM amplified mode) from the two Juniper routers (ptx10001) are amplified (not shown in (b)) and then connected to a 16x4 filter, followed by a 12-degree ROAMD (RD-12RS) equipped with booster and pre-amplifier (AM_S23L). One SSMF spool of 100km between node F8_41and F8_45. Rather than listing all network operation results, Fig. 4(c) highlights the GNPy-based per- formance estimation for the example service path. For reference, the OSNR including nonlin- ear interference (NLI) under full C-band loading is shown alongside the measured OSNR for sin- gle-channel transmissions at 191.5 , 193.1, and 196 THz. The performance metrics of the two co- herent transceivers are reported in Fig. 4(d). Conclusions We proposed an Agentic AI orchestrated collab- oratively multi‑MCP server architecture enabling full and efficient SDN‑based automation. Experi- mental validation on a real IPoDWDM testbed demonstrates the feasibility to streamline E2E service lifecycle operation and management across physical, optical and IP layers. Acknowledgment This work has received funding from the German Federal Min- istry of Research, Technology, and Space (BMFTR) project SUSTAINET-Advance, Grant 16KIS2271K, in the framework of the CELTIC-NEXT project id C2024/3-3. References [1] Model Context Protocol (MCP), https://modelcontextpro- tocol.io/ (2024). [2] Ll. Gifre, R. Vilalta, R. Munoz, “Vision 2030: Autono- mous Network Control in an All Optical World”, Tu2C.1, OFC2026. [3] Xiaonan Xu, Haoshuo Chen, Sarvesh Bidkar, Jesse E. Simsarian, Ells Burrows, Roland Ryf, Mikael Mazur, Ni- colas K. Fontaine, and David T. Neilson, “Distributed and Dynamic AI Agent Collaboration over Optical Transport for Network Orchestration and Monitoring”, Tu2C.2, OFC2026. [4] S. Cruzes, “Telemetry and Agentic AI: Foundations for Optical Network Automation,” TechRxiv, Dec. 2025. DOI:10.36227/techrxiv.176539548.84734375/v [5] Q. Zhuge, X. Liu, Y. Zhang, and Q. Qiu, “Digital Twin and AI Agent for Autonomous Optical Networks”, in Pro- ceedings of the Optical Fiber Communication Confer- ence (OFC 2025), San Francisco, CA, USA, 2025, paper W4G.1. DOI: 10.1364/OFC.2025.W4G.1 [6] X. Liu, Q. Qiu, Y. Zhang, Y. Cheng, L. Yi, W. Hu, and Q. Zhuge, “First Field Trial of LLM‑Powered AI Agent for Lifecycle Management of Autonomous Optical Net- works,” presented at Optical Fiber Communication Con- ference (OFC 2025), San Francisco, CA, USA, 2025. paper Th1A.2. DOI: 10.1364/OFC.2025.Th1A.2 [7] Y. Zhang, Y. Song, S. Li, Y. Shi, S. Shen, X. Tang, M. Zhang, and D. Wang, “Generative AI‑Driven Hierarchical Multi‑Agent Framework for Zero‑Touch Optical Net- works,” DOI: 10.48550/arXiv.2510.05625 [8] Óscar González de Dios, Pablo Armingol Robles, Liesbeth Roelens, Alejandro Muñiz-Da-Costa, Ignacio de Miguel, Ramón J. Durán Barroso, and Juan Pedro Fernández-Palacios., “Automation of multi-layer multi domain transport networks and the role of AI,” J. Opt. Commun. Netw., vol. 17, no. 2, p. A124–A134, Feb. 2025, [Online]. DOI: doi.org/10.1364/JOCN.537463 [9] Hussein Zaid, Pooyan Safari, Behnam Shariati, Aydin Jafari, Mihail Balanici, Johannes Karl Fischer., “Multi- Agent Design for LLM-assisted Network Management,” Optical Fiber Communication Conference (OFC), San Diego, CA, USA, Feb. 2025. [10] Oneline: The EDC Connector: Driving Data Sovereignty - Think-it [11] Angela Mitrovska, Hussein Zaid, Behnam Shariati, Pooyan Safari, Johannes Karl Fischer, Ronald Freund, “Demonstration of Customer-Owned and Carrier-Con- trolled Coherent Pluggables using Ownership-Aware YANG Models”, OFC2026 [12] Online available : https://gnpy.readthedocs.io/en/mas- ter/json.html

Instant research discovery