More efficient LLMs and advances in Agentic AI are enabling a new network automation paradigm. In this new model, multiple AI agents representing specific network areas or performance measures collaborate to operate a network autonomously. Natural language interfaces enable easy interaction with telecoms employees.
Communications service providers (CSPs) are involved in a long-term effort to automate their network operations. Their goal is for networks to diagnose and correct both mechanical faults and service problems with minimal human intervention.
AI in general – and generative AI in particular – consume huge amounts of data centre resources as well as the electricity and water used to support them. Recent advances in LLMs – of which DeepSeek is the most well-known – have improved the efficiency of the inference phase: it takes fewer resources to respond to a prompt, which means that the model responds more cheaply and in less time, making it more useful for real-time applications.
This, in turn, has led to a multi-agent model: instead of a single AI instance trying to manage all aspects of a network’s operation, the new model relies on multiple agents, each responsible for a particular concern, for example SLA compliance, capacity expansion, or service provisioning.
These agents communicate with each other to reach a collaborative solution to a given issue. In essence, each agent focusses on closed-loop automation in its own domain while collaborating with other agents on cross-domain issues.
The benefits of agentic collaboration for telecoms
This type of collaboration has benefits for both the humans who are in charge of the network as well as the system in which they operate. Because the agents can conduct their collaboration in natural language, human employees can easily follow the discussion, intervene when necessary, and make operational decisions when warranted.
US Tariffs are shifting - will you react or anticipate?
Don’t let policy changes catch you off guard. Stay proactive with real-time data and expert analysis.
By GlobalDataJust as important for a highly regulated industry like telecoms, this natural language record of the decision-making process provides reliable auditability: when a telecoms employee, customer, or regulator needs to know why a certain network performance decision was made, the natural language record provides an easy-to-understand answer.
How agents learn
AI agents’ natural language abilities also enable them to learn from internal manuals, troubleshooting procedures, incident reports, and so on. More importantly, it can learn from human experts. This can take the form of human feedback to an automated change or even of “coaching,” where the agent suggests a sequence of actions and the human adjusts it.
Work on applying this capability to network operations in China Guangdong indicates that the agent can learn not only from this coaching, but from its own success or failure in addressing problems. The agent thus increasingly becomes capable of addressing new problems and uncertain scenarios independently, making its interaction with human personnel closer to a relationship of equals.
Assigning an agent to each domain and area of emphasis also helps the CSP integrate agents into its operations. To a certain extent, this agent structure can be modified to match the organisational structure of the operations group, so that processes and operational models can remain largely constant even as the network substantially increases its automation.
The multi-agent approach
On the technical side, the multi-agent approach can isolate the impact of failures, whether due to a bad upgrade or a technical issue. Since the outage will only affect the agent that is directly tied to its domain, the rest of the system will function as usual. This is an advantage over a single AI instance that runs the whole network, since a failure in one domain makes it more likely that other, unrelated systems would be affected.
A multi-agent model like this does require certain technical transformations. Fortunately, CSPs are already working on the necessary components. Chief among them is a digital twin of network and service operations. The digital twin collects information from all parts of the network, transforms and stores it, supplies it to other systems, and models the interaction of the physical network and service performance across the network. In order to collaborate effectively, AI agents must have access to a single source of truth about how the network is serving its users.
To enable multi-agent collaboration, there must also be an agent architecture to provide common tools and data to the agents, as well as to handle interaction with the digital twin.
The architecture must also specify rules for process combination and conflict resolution. It must be able to learn from application experience data and incorporate input from every new generation of intelligent hardware. Structuring network automation around multiple ‘virtual colleagues’ fuelled by a common digital twin should ease the progression toward fully autonomous networks.
