Autonomous Network Operations Using Agentic AI

british_telecom · Telecommunications · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Telecommunications

BT and AWS are applying agentic AI to autonomous network operations, aiming to reduce the cost and time required to operate complex telecom networks.

Constraints:

large and complex 5G network operations
multiple network lifecycle phases
network state and operational knowledge must be available to agents
enterprise-grade security and reliability requirements

Architecture

BT's autonomous network operations architecture combines network telemetry, graph topology, vectorized operations knowledge, SageMaker models, Amazon Bedrock/AgentCore agents, Neptune graph analytics, RDS/S3 storage, and ticketing-system integrations. The system turns alarms, anomalies, incidents, changes, topology, KPIs, and operational documents into root-cause and service-impact analysis, with domain-specific community agents correlating alarms across network communities and domains.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  network-telemetry-ingest["network-telemetry-ingest"]
  ai-data-product-layer["ai-data-product-layer"]
  agentic-semantic-layer["agentic-semantic-layer"]
  network-digital-twin["network-digital-twin"]
  operations-knowledge-vector-store["operations-knowledge-vector-store"]
  msk-alarm-stream["msk-alarm-stream"]
  agentcore-runtime["agentcore-runtime"]
  neptune-topology-graph["neptune-topology-graph"]
  domain-community-agents["domain-community-agents"]
  service-impact-analysis["service-impact-analysis"]
  rds-alarm-store["rds-alarm-store"]
  s3-service-impact-store["s3-service-impact-store"]
  ticketing-integration["ticketing-integration"]
  network-telemetry-ingest --> ai-data-product-layer
  network-telemetry-ingest --> msk-alarm-stream
  ai-data-product-layer --> agentic-semantic-layer
  ai-data-product-layer --> network-digital-twin
  agentic-semantic-layer --> operations-knowledge-vector-store
  network-digital-twin --> neptune-topology-graph
  network-digital-twin --> domain-community-agents
  operations-knowledge-vector-store --> domain-community-agents
  msk-alarm-stream --> agentcore-runtime
  agentcore-runtime --> domain-community-agents
  agentcore-runtime --> neptune-topology-graph
  agentcore-runtime --> operations-knowledge-vector-store
  neptune-topology-graph --> domain-community-agents
  domain-community-agents --> service-impact-analysis
  domain-community-agents --> ticketing-integration
  service-impact-analysis --> rds-alarm-store
  service-impact-analysis --> s3-service-impact-store
  service-impact-analysis --> ticketing-integration

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`network-telemetry-ingest`	collects alarms, anomalies, events, incidents, topology, changes, KPIs, and documentation inputs	`ai-data-product-layer`, `msk-alarm-stream`
`ai-data-product-layer`	curates network data and builds features for ML and agentic applications	`agentic-semantic-layer`, `network-digital-twin`
`agentic-semantic-layer`	defines KPI, alarm, and correlation meanings once for runtime reuse	`ai-data-product-layer`, `operations-knowledge-vector-store`
`network-digital-twin`	combines topology, service models, performance, alarms, and historical/current network views	`neptune-topology-graph`, `domain-community-agents`
`operations-knowledge-vector-store`	makes unstructured operations documentation and RCA knowledge available to agents	`domain-community-agents`
`msk-alarm-stream`	streams alarms and anomalies into the agentic RCA workflow	`agentcore-runtime`
`agentcore-runtime`	runs Bedrock/AgentCore primitives for RCA and service-impact agents	`domain-community-agents`, `neptune-topology-graph`, `operations-knowledge-vector-store`
`neptune-topology-graph`	represents network connectivity and alarm clusters for graph analytics	`domain-community-agents`
`domain-community-agents`	domain-specific agents correlate alarms inside and across network communities to identify root causes	`service-impact-analysis`, `ticketing-integration`
`service-impact-analysis`	maps RCA outputs to affected customers, services, and customer-experience metrics	`rds-alarm-store`, `s3-service-impact-store`, `ticketing-integration`
`rds-alarm-store`	stores alarm records used by the RCA workflow	`service-impact-analysis`
`s3-service-impact-store`	stores service impact metrics	`service-impact-analysis`
`ticketing-integration`	creates, updates, and deletes tickets based on RCA and impact analysis results	`domain-community-agents`, `service-impact-analysis`

Data flow

Raw network data such as counters, alarms, topology, configs, incidents, changes, knowledge sources, KPIs, and customer-experience metrics is ingested into data products, semantic layers, graph databases, vector stores, and a network digital twin. Alarms and anomalies flow via MSK to AgentCore runtime agents, which use Neptune topology and operations knowledge bases for root-cause analysis; service-impact analysis combines RCA output with customer metrics and writes alarms to RDS, service impact metrics to S3, and ticket updates to trouble-ticketing systems.

3 supporting quotes

“counters, alarms, topology, config, incidents, changes, knowledge sources, all of these acts as the raw data that is required”transcript, implied

“alarms and anomalies coming through via MSK from the left, and then the agents that are deployed on runtime use Amazon Neptune”transcript, stated

“the alarms are stored on Amazon RDS and the service impact metrics are stored on Amazon S3.”transcript, stated

Why this architecture

BT's network scale makes manual operations and static rule automation insufficient. Containerized 5G infrastructure emits huge telemetry volumes, topology is connected and graph-shaped, operations knowledge is partly unstructured, and service impact depends on customer metrics. The architecture therefore combines curated data products, graph topology, vectorized knowledge, and domain/community agents rather than a single generic agent.

3 supporting quotes

“Supervised ML models do not give enough performance today, because of lack of availability of good training data.”transcript, stated

“network topology is what underpins this and it's often incomplete and inaccurate.”transcript, stated

“Networks are designed in a way to keep them resilient as well. So when they are designed, they're designed with a blast radius in mind.”transcript, stated

Models used: Amazon Bedrock base models · small language models · SageMaker anomaly detection models · SageMaker impact-prediction models

Scale

BT's mobile network has about 30 million provisioned subscribers and 22.5 million daily mobile users.
“we have about 30 million provisioned subscribers and every day about 22.5 million mobile users use our mobile network”transcript, stated
BT aims for 99% UK population coverage by 5G standalone by late 2029 or early 2030.
“by end of 2029, early 2030 to have 99% of UK population covered by 5G standalone.”transcript, stated
The network includes 20,000 macro sites and growing.
“you see 20,000 macro sites. So these are the big antennas and the poles that you see in the streets.”transcript, stated
The radio network monitors 4,000 KPIs for each tower.
“the radio itself, we monitor 4,000 KPIs for each tower.”transcript, stated
BT cited about 12,000 events a day and hundreds of changes per day.
“per day we have, you know, some of these stats, you know we have for example 12,000 events a day. We do hundreds of changes in a day”transcript, stated
The mobile network sees 11,000 changes per week.
“per week we do in our mobile network 11,000 changes and majority of the time we're successful.”transcript, stated

Alternatives considered (roads not taken)

Manual network operations at current scale

Rejected because: BT describes current operations as not where they need to be for a containerized 5G network with huge telemetry and change volume.

“our operation is not really where it needs to be in order to deal with the ambition that we have”transcript, stated

Rule-based automation as the main RCA mechanism

Rejected because: The RCA section says rule-based automation has heavy cost.

“There is heavy cost in rule-based automation.”transcript, stated

Supervised ML-only RCA

Rejected because: BT says supervised ML models lack enough good training data for the RCA task.

“Supervised ML models do not give enough performance today, because of lack of availability of good training data.”transcript, stated

One generic network agent for all topology

Rejected because: BT uses domain-specific community agents because network domains and blast-radius communities need local and cross-community correlation.

“we called it domain-specific community agents.”transcript, stated

“agents that are operating within these communities collaborate with one another and then share knowledge with other agents”transcript, stated

Patterns

`multi-agent-orchestration` (orchestration)

The architecture is described as a complex agentic system with agent primitives and a family of agents for autonomous networks.

“And think of AgentCore as a suite of primitives to help build complex agentic systems that are scalable, that are enterprise grade with security and reliability that's required of a large complex agentic application.”transcript, stated

`rag-retrieval` (retrieval)

Operational documentation and historical incident knowledge are vectorized so agents can retrieve relevant context for current network conditions.

“This vectorized knowledge base enables LLM-based agents to access relevant context when reasoning about current network conditions.”summary, stated

Gotchas (what bit them)

The autonomous-network vision spans more lifecycle phases than the initial implementation covers.

Root cause: Planning, configuration, operations, and service fulfillment are broader than a first operations-focused agentic deployment.

Resolution: The transcript frames expansion into planning and configuration as future work over time.

“And we, over time we expect agentic AI to expand to these other phases of planning and configuration as well.”transcript, stated

Evaluation

The summary describes versioning, evaluation, feedback integration, and continuous improvement as part of the foundation for later agentic components.

“The discipline of versioning, evaluation, feedback integration, and continuous improvement establishes patterns that extend to the LLM-based agentic components built atop this foundation.”summary, stated

Outcomes

The transcript describes the strategic direction and architecture, but not quantified production outcomes for the agentic system.