Intent-Based Networking: Agentic Network Automation

Solution Overview

Telecommunications and media are among the industries that lead AI adoption, alongside software and fintech. Yet the gap between leaders and everyone else keeps widening: companies at the frontier of AI capability generate roughly double the revenue growth of their peers, while nearly three-quarters of all companies have yet to show any tangible value from their AI investments. Telecom races ahead on one front in particular — it has the highest rate of agentic AI adoption of any industry. The gap isn't the model. It's everything underneath that agents need to act on live network data:

• Data pipelines

• Vector search

• Embeddings

• Short-term and long-term memory

• Real-time processing

Agentic network automation closes that gap. Instead of engineers translating requests into device configurations by hand, autonomous agents interpret goals, plan changes, act on the network, and verify the result. IBN is the clearest expression of this pattern. You state a network outcome in plain business terms, and the system delivers and defends it without further intervention.

With IBN, an operator's customer describes what they need, not how to build it. For example: Open a new flagship store. Give POS traffic priority, keep guest WiFi strictly separated, add a camera uplink, and hold POS latency under 40 ms.

A central agent translates that intent into network policies, provisions the services, and monitors them continuously. When the network drifts from the promise, the agent diagnoses the cause and applies a fix on its own.

MongoDB Atlas serves as the operational data layer driving this workflow. It consolidates every resource the agent needs in one database, delivering real-time queries across the current network state, customer agreements, event histories, and operator insights. The AI layer connects directly to the data, eliminating message bus, cache, or ETL pipelines. The Reference Architectures section shows how this framework works in detail.

Reference Architectures

The solution runs on a single agent backed by MongoDB Atlas. Figure 1 shows the main components:

• A ReAct-based orchestrator

• A two-stage semantic router

• A set of specialized MCP microservices

• Atlas collections storing the network's data and the agent's memory

Every service reads from and writes to Atlas, eliminating the need to maintain message bus, cache, or ETL pipelines. This foundation is extensible: the same orchestrator, router, and memory layer take on new domains by adding MCP services and collections. This framework sets the stage for future use cases such as a Digital Network Twin for what-if capacity planning.

Figure 1. The orchestrator routes each query through MongoDB Atlas, which stores the service catalog, the IBN collections, and the agent’s memory.

Data Model Approach

IBN works with different data shapes, and MongoDB Atlas holds them all in one place. Each of the following collections maps to a part of the workflow:

• ibn_intents: Stores the parsed intent and its lifecycle state. It contains every field specified in the request, such as the latency ceiling or the segmentation policy.

• ibn_sites: Contains the network sites with 2dsphere-indexed coordinates for geospatial lookups.

• ibn_resources: Contains the network resources available at each site.

• ibn_policy_snapshots: Stores immutable plan snapshots that preserve the full planning history.

• ibn_telemetry. Stores metric samples in a time series collection.

• ibn_compliance_events: Stores the record of every violation and recovery.

• ibn_knowledge_chunks: Stores past incidents, runbooks, and templates, auto-embedded with Voyage AI for vector search.

The agent's memory also lives in MongoDB, stored in dedicated collections alongside the network data:

• agent_workstreams: Stores short-term context for the current work thread.

• agent_memories: Extracts long-term facts when a workstream closes, vector-indexed for recall across sessions.

• user_preferences: Stores instructions the engineer teaches the agent.

A single Atlas Vector Search index makes the four-dimensional diagnosis query possible. With Atlas auto-embedding, you point the index at a text field and Atlas generates and stores the embeddings for you. You don’t need a separate embedding pipeline or service to run. That same index pairs the auto-embedded text field with structured, time, and geospatial filters. As a result, a $vectorSearch stage does the work of several query engines:

{
   "fields": [
      {
         "type": "autoEmbed",
         "modality": "text",
         "path": "text",
         "quantization": "float",
         "model": "voyage-4"
      },
      {
         "type": "filter",
         "path": "kind"
      },
      {
         "type": "filter",
         "path": "segment"
      },
      {
         "type": "filter",
         "path": "market"
      },
      {
         "type": "filter",
         "path": "plan_id"
      },
      {
         "type": "filter",
         "path": "ts"
      },
      {
         "type": "filter",
         "path": "lng"
      },
      {
         "type": "filter",
         "path": "lat"
      }
   ]
}

Build the Solution

The full demo is available on this GitHub repository. Clone the repository, then follow these steps.

export OPENAI_API_KEY="<your openai api token>"
export MONGODB_URI="<your mdb connection string>"
export VOYAGE_API_KEY="<your voyage api token>"

brew install python@3.13
export PATH="$(brew --prefix)/opt/python@3.13/libexec/bin:$PATH"
python -m venv <dir>
source <dir>/bin/activate
cd agentic-mcp-demo
pip install -r requirements.txt

./bin/start.sh

-I'm opening a new Alpenmarkt store at Marienplatz Munich. POS priority, guest WiFi strictly separated, camera uplink, online by 18:00, max 40ms POS latency, 99.95% availability
-feasibility check
-propose and activate
-inject morning rush
-diagnose violation
-apply runbook

Key Learnings

• Unify your data in one store: Keep intent records, geospatial sites, time series telemetry, and vector-indexed knowledge in a single MongoDB Atlas database, queryable with one driver and one pipeline.

• Retrieve across dimensions in one query: Combine vector similarity, structured filters, a time window, and geospatial bounds in a single Atlas Vector Search stage, with no application-side orchestration.

• Stream real-time changes: Use MongoDB Change Streams to push intent activations, violations, and recoveries to dashboards without polling.

• Give your agent a memory: Store short-term context, long-term facts, and user preferences as collections, so the agent improves with use rather than retraining.

• Automate the full intent lifecycle: Let one agent parse, plan, activate, assure, and remediate network intents end to end, grounded in live data.

Authors

• Benjamin Lorenz, MongoDB

• Aditya Vikram Roy, MongoDB

• Diego Canales, MongoDB

Learn More

• AI-Driven Media Personalization
• AI-Powered Chatbot for Network Management
• Gen AI-Powered Video Summarization