The telecoms industry will need to address and solve issues around interoperability, security, governance and data before Agentic AI can deliver autonomous networks at commercial scale, according to a new report from the Next Generation Mobile Networks Alliance.
The NGMN report, “Network Automation and Autonomy Phase III: Agentic AI for Autonomous Mobile Networks“, argues that by acting to co-ordinate domain-level automations, Agentic AI could become a key enabler of the transition towards operational autonomy, in which networks can make decisions and execute actions with minimal human intervention.
However, the report warns that the industry’s current ecosystem is highly fragmented with vendors using different data models and terminology, making it difficult for agents to develop a consistent understanding of network conditions across domains. As a result the paper calls for greater alignment between organisations including 3GPP, TM Forum, ETSI, O-RAN, IETF, W3C and BBF, including common information models, ontologies and semantic mappings.
These cross-organisation challenges fall into eight areas that NGMN says can be addressed by its own reference framework :
A) Fragmentation across organisations
B) Limited focus on multi-vendor interoperability and agent coordination across domains
C) Lack of standards for agent knowledge- and context-sharing mechanisms
D) Incomplete information modelling for networks
E) Lack of end-to-end security, identity and trust for autonomous functions
F) Limited focus on the human-machine interface and operational integration
G) Lack of operator-friendly governance and lifecycle tools
H) Under-addressed economic and organisational readiness
As well as this ecosystem alignment, the paper calls for interoperability and the development of the supporting systems required to make autonomous AI trustworthy and controllable. That means suporting interoperable agent communication, shared semantics, secure identity, auditable execution, digital-twin validation, lifecycle assurance and practical human-in-the-loop or human-on-the-loop controls.
The paper also highlights internal organisational challenges facing operators, stating, “While technology groups focus on architecture, few have formal guidance on the required telco organisational changes.”
Cross-domain supervision
According to the paper, most existing network automation is based on predefined rules and fixed workflows. While this works effectively for repetitive tasks, it becomes harder to manage when network conditions change or when an operational problem crosses multiple network domains.
As a result, operators can have highly automated individual domains but still require engineers to investigate problems and coordinate actions across RAN, core, transport and cloud domains. NGMN says the industry needs to move towards end-to-end, controlled closed-loop automation driven by desired service outcomes.
And it is here that the role of Agentic AI could help by delivering the ability to discover and invoke existing management systems, controllers and automation tools, while dividing complex objectives into smaller tasks.
For example, a supervisory agent could coordinate specialist agents that each cover RAN, core, transport, cloud and security to diagnose a service problem, assess potential corrective actions, resolve conflicts and verify that service has been restored. The NGMN paper uses three scenarios to demonstrate the approach:fault management, service assurance and RAN optimisation. In each case, agents could divide activities such as data collection, diagnosis, impact assessment, solution generation, validation and controlled execution.
In the RAN optimisation example illustrated below, the paper says:
“The workflow is triggered by changes in telemetry data or an input from a service assurance agent (1). On the service operations layer, a cross-domain agent is delegating intent and allocating tasks to a RAN-specific optimisation agent also sharing policies (2). Within the RAN domain, an optimisation agent in the resource operations layer acts as a supervisor agent and delegates a task to a sub-agent residing in the infrastructure layer (3). Subagents may query and share information directly between them (4). All agents can invoke APIs, tools, knowledge graphs and digital twins, which may logically operate on different levels like service operations, resource operations and in the infrastructure layer. Where RAN optimisation actions run into conflicts with e.g. network-wide energy management, the cross-domain agent guides conflict resolution with the relevant energy management agent (5) and may arbitrate trade-offs between performance and energy consumption goals. Existing and possibly vendor-specific proven RAN automation and management components may continue to support deterministic and reliable actions through exposing their capabilities via APIs or agentic tool interfaces to any L4-capable RAN optimisation agent, as illustrated in the figure.”
Setting up the guardrails
The report says operators need to know exactly what an autonomous agent is permitted to do and how its behaviour can be monitored. It calls for a telecom-grade Zero-Trust Agent Ecosystem, incorporating secure agent identity, authentication, authorisation, policy enforcement, audit logging, runtime monitoring and the ability to revoke or quarantine agents.
Laurent Leboucher, Chairman of the NGMN Alliance Board and Orange Group CTO and EVP Networks, said: “The rapid arrival of Agentic AI has introduced the prospect of autonomous systems able to reason, plan, collaborate and execute actions within and across multiple operational domains.
“However, the complexity of networks and high expectations surrounding customer experience means general-purpose Agentic AI technologies must be used carefully with the right harness and guardrails to bring humans in the loop when needed. This is a learning process in itself where operators will have to take controlled risks.”
