Artificial intelligence is changing the way telecom operators think about network management. AI agents can monitor networks, understand alarms, analyse performance data, find problems and, in some cases, take action without waiting for a human operator.
This could make telecom networks faster, more efficient and more autonomous.
But there is another side to the AI revolution.
As telecom companies start buying pre-built AI agent libraries in telecom, they could create a new type of vendor dependence. The concern is not necessarily the AI model itself. Operators can increasingly use different large language models and AI technologies. The bigger concern is the platform that controls how those AI agents communicate, discover each other, access network data, and take action.
A recent Telecoms Tech News analysis describes this as a possible new OSS trap for telecom operators. The argument is that agent libraries could become similar to the old telecom operational support systems (OSS) that operators found difficult and expensive to replace.
The issue matters because telecom companies are now moving toward autonomous networks, where AI agents may eventually make decisions across radio access networks (RAN), transport, core networks and OSS/BSS systems.
If the industry chooses closed agent platforms too early, today’s shortcut to automation could become tomorrow’s integration problem.
What Are AI Agent Libraries in Telecom?
An AI agent is software that can understand a task, use available information and tools, make decisions and take actions within defined limits.
In telecom, an agent could be designed for a specific job.
For example, one agent might monitor network alarms. Another could analyse key performance indicators (KPIs). A third could investigate an anomaly. Another could recommend or perform a corrective action.
Instead of one general AI system trying to do everything, operators can use groups of specialised agents.
AI agent libraries package many of these specialised capabilities together.
The idea is attractive because telecom operators do not have to build every agent from scratch. Vendors can provide pre-built agents that already understand telecom concepts, network data and operational processes.
Nokia, Samsung, Amdocs and Ericsson are among the major vendors developing platforms and agent-based capabilities for autonomous network operations.
This could reduce the time needed to move from an AI experiment to a real operational system.
But it also creates an important question:
Who controls the platform through which all these agents work?
Why Telecom Operators Want AI Agents
Telecom networks are extremely complex.
A modern operator may manage mobile networks, fibre, cloud infrastructure, data centres, 5G core systems, enterprise services and thousands of network elements.
Human teams cannot manually investigate every alarm or optimisation opportunity.
AI agents can potentially help by continuously watching network conditions and responding to problems.
For example, imagine a mobile cell showing unusual uplink behaviour.
A traditional process may involve:
- An alarm appears.
- A network engineer investigates it.
- The engineer checks KPIs.
- Different systems are opened to understand the problem.
- A possible cause is identified.
- A solution is tested.
- A change is approved.
- The change is implemented.
- Engineers monitor the result.
An agentic system could connect many of these steps.
One agent could detect the anomaly. Another could analyse the relevant KPIs. A reasoning agent could identify possible causes. Another agent could recommend an action, while a policy system checks whether that action is safe.
If approved, the action could be executed automatically and monitored.
This is the basic idea behind the move toward autonomous networks.
Ericsson, for example, describes autonomous network operations as using intent-based control, agentic orchestration, real-time observability and closed-loop AI. Its approach includes specialised agents working across RAN, transport, core and OSS/BSS domains, with guardrails and approval levels for high-impact actions.
The Real Risk May Not Be the AI Model
One of the most important points in the Telecoms Tech News analysis is that the underlying AI model may not be the biggest lock-in risk.
Operators can potentially change models.
They could use different large language models depending on cost, performance, privacy, or business requirements.
The harder component to replace could be the agent control layer.
This layer can determine:
- Which agents are available
- How agents discover each other
- Which agent receives a task
- What data an agent can access
- What actions an agent is allowed to take
- Which policies apply
- How decisions are recorded
- How agents communicate
- How network context is represented
Once an operator builds its processes around one vendor’s registry, ontology, policies and agent runtime, moving to another platform could become much harder.
The Telecoms Tech News analysis compares this possibility with historical OSS lock-in, where operators sometimes became deeply dependent on vendor-specific inventory, assurance and order-management systems.
The difference is that AI agents may accumulate not only configuration data but also operational decision context.
That could make future migration even more complicated.
Nokia, Samsung, Amdocs and Ericsson Are Building Different Approaches
The current market already shows several different approaches to agentic telecom operations.
Nokia’s Agent Library
Nokia is developing a Gemini-backed Agent Library with specialised agents for tasks such as routing, event triage, KPI selection, anomaly reasoning and action reasoning.
According to Telecoms Tech News, these agents are built using Google’s Agent Development Kit and are connected to Nokia’s network assurance and autonomous-network products.
The advantage is clear: operators can start with pre-built telecom capabilities rather than building every agent themselves.
The potential challenge is deciding how portable those agents will be outside the Nokia environment.
Samsung’s Agent Fabric
Samsung has placed its AI Agent Fabric at the centre of CognitiV Network Operations Suite.
Samsung describes Agent Fabric as a decision-intelligence layer that allows multiple specialised AI agents to work together instead of operating as isolated tools. The company says the system can support activities across the network lifecycle, including planning, maintenance, troubleshooting and optimisation.
This is important because the agent fabric becomes more than a collection of individual AI tools.
It becomes the coordination layer.
And coordination is exactly where interoperability becomes important.
Amdocs aOS and Cognitive Core
Amdocs has introduced aOS, which it describes as an agentic operating system designed specifically for telecommunications.
The company says aOS can operate on top of existing BSS/OSS environments and uses its Cognitive Core to provide telecom-specific agent libraries and intelligence.
Amdocs’ published architecture describes pre-built agent and sub-agent libraries across areas such as sales, care, monetisation, service delivery and network management, alongside telecom context, ontology, data, security, governance and observability.
That broad scope shows why agent platforms could become strategically important.
If an operator uses agents across many business and network functions, replacing the platform becomes a much bigger project.
Ericsson’s EIAP and cApps
Ericsson is also expanding its Intelligent Automation Platform (EIAP).
In June 2026, Ericsson announced that EIAP was being expanded from RAN automation into core network automation. The company introduced cApps for core networks and described how AI-driven applications could support autonomous operations.
Ericsson says EIAP is designed as an open, multi-vendor automation platform and is already used by CSPs including AT&T, Swisscom, Telstra and Vodafone.
The company’s wider AI-native strategy also calls for specialised agents to work across multiple network domains while operating within shared policies and safety boundaries.
These different approaches show that the industry is moving quickly.
But they also raise the interoperability question.
The Interoperability Problem
The biggest challenge may be simple:
Can an AI agent created for one telecom platform work easily on another?
In theory, standards such as APIs, MCP, and agent-to-agent communication can help.
But basic communication is not the same as full interoperability.
An agent needs to understand telecom-specific information.
For example, it needs to know what a particular network event means, what a KPI represents, what network element is involved and what action is safe.
Different platforms may represent this information differently.
The Telecoms Tech News analysis argues that without a shared operational contract for telecom agents, vendors could effectively create their own private agent languages.
That could create a familiar problem for telecom operators.
Instead of integrating different OSS platforms, they may eventually find themselves integrating different agent fabrics.
Why MCP and A2A Are Not the Complete Answer
Two technologies frequently discussed in agentic AI are Model Context Protocol (MCP) and Agent2Agent (A2A).
They can help AI systems interact with tools and communicate with other agents.
However, telecom networks require more than basic agent communication.
A production telecom agent may need:
- Identity and authentication
- Network topology
- Service context
- Inventory information
- Policy controls
- Risk classification
- Human approval
- Audit trails
- Real-time monitoring
- Rollback mechanisms
- Telecom-specific data models
That is why simply supporting MCP or A2A does not automatically guarantee vendor-neutral interoperability.
The Telecoms Tech News analysis makes the same distinction, arguing that these protocols should be treated as a foundation rather than the entire interoperability solution.
Why L4 Autonomous Networks Make This More Important
The industry is increasingly discussing Level 4 autonomous networks, where systems can perform many operational tasks automatically while humans remain involved in higher-risk situations.
This is a major reason vendors are pushing agentic AI now.
Ericsson says Level 4 autonomy requires intent-based operations, agentic orchestration, observability, closed-loop automation and strong guardrails.
Samsung similarly describes Agent Fabric as part of its path toward highly autonomous networks and L4/L5 operations.
But the closer networks move toward autonomous operation, the more important the control layer becomes.
If an AI agent is only generating a recommendation, switching platforms may be relatively easy.
If thousands of agents are making or coordinating production network changes, the platform becomes much more deeply embedded.
How Telecom Operators Can Avoid a New OSS Trap
Operators do not necessarily need to avoid vendor AI agents.
Pre-built agents can provide real value and speed up automation.
The important issue is where to draw the architectural boundary.
Telecom companies should consider several principles before making long-term commitments.
1. Separate the Agent Bus From the Agent Library
Operators should avoid automatically treating the vendor’s agent library and its underlying orchestration platform as one inseparable product.
The control layer should ideally remain replaceable.
2. Demand Vendor-Neutral Agent Registration
An agent should be discoverable through a common registry rather than only through a vendor’s private catalogue.
That could make it easier to introduce agents from multiple suppliers.
3. Make Policies Portable
Policies, approval rules, and audit information should be exportable.
An operator should not lose its governance model simply because it changes platforms.
4. Test Real Multi-Vendor Agents
It is easy to demonstrate interoperability in a presentation.
A better test is to run agents from different vendors against real operational workflows.
Operators should ask vendors to demonstrate this before signing long-term agreements.
5. Protect the Data and Ontology Layer
The meaning of telecom data is extremely important.
AI agents need consistent information about network elements, services, customers, topology, and performance.
Operators should avoid allowing one vendor to become the only owner of that operational meaning.
6. Keep Humans in High-Risk Loops
Not every network action should be fully autonomous.
Actions that could affect large numbers of customers, critical infrastructure, or regulatory obligations should have appropriate approval and rollback mechanisms.
Ericsson’s own AI-native guidance highlights policy constraints, approval tiers, and continuous KPI verification as important safety mechanisms.
AI Agents Could Still Be a Major Opportunity
The OSS lock-in risk should not hide the potential benefits.
AI agents could fundamentally improve telecom operations.
They can help reduce repetitive manual work, speed up incident investigation, improve network visibility and support faster optimisation.
Research into multi-agent network operations is also exploring practical uses such as packet-loss analysis, node isolation, vendor knowledge searches and automated remediation.
The goal should therefore not be to stop using AI agents.
The better goal is to make the agent ecosystem open enough that operators remain in control.
The Future of AI Agent Libraries in Telecom
The telecom industry has already learned that infrastructure decisions can last for decades.
That lesson is especially important now.
AI technology changes extremely quickly. Today’s most powerful model may be replaced by a better, cheaper or more specialised model in a short period.
But the systems built around that model can last much longer.
That is why the strategic question for telecom operators is not simply:
“Which AI agent library is the best?”
It should also be:
“Can we replace the agent library without rebuilding our entire operating model?”
The answer will depend on interoperability, open standards, portable policies, shared telecom semantics and vendor-neutral orchestration.
If these principles develop alongside AI agents, operators could get the benefits of automation without creating another generation of proprietary OSS dependencies.
If they do not, the industry could solve one integration problem only to create another.
