Agentic AI Could Enable Level 4 Autonomy, but the Ecosystem Needs to Catch Up

On August 11, 2026, the Next Generation Mobile Networks Alliance (NGMN) published a report examining how Agentic Artificial Intelligence (AI) can enable Mobile Network Operators (MNOs) to progress from Level 2/3 automation to Level 4 autonomous networks. This means moving from networks that automate specific tasks or make decisions with human oversight toward systems that can make and execute decisions across multiple network domains with minimal human intervention. Under the TM Forum framework, Level 4 is defined as a “highly autonomous” network capable of making decisions in complex, cross-domain environments through predictive analysis or active closed-loop management. MNOs have already deployed Machine Learning (ML), closed-loop optimization, and domain-specific controllers for Radio Access Network (RAN) optimization, fault management, service assurance, and energy management. The challenge is that these capabilities remain siloed within individual network domains, making cross-domain operations difficult or, in some cases, impossible to automate.

NGMN argues that Agentic AI could close this gap by allowing AI agents to interpret intent, plan actions, delegate tasks to other agents, invoke existing network functions, and validate results. The report emphasizes that agents will need to complement existing automation investments and identifies requirements including interoperable agent communication, shared data and context, secure agent identity and authorization, policy enforcement, observability, auditability, lifecycle management, and human oversight.

NGMN also highlights fragmentation across standards organizations, open-source communities, and vendors as a potential barrier to adoption. Work on Agentic AI for network automation is already underway across standards organizations and open-source communities, such as The 3rd Generation Partnership Project (3GPP), European Telecommunications Standards Institute (ETSI), TM Forum, Internet Engineering Task Force (IETF), World Wide Web Consortium (W3C), O-RAN Alliance, Broadband Forum (BBF), and the Linux Foundation. However, NGMN warns that these differences across initiatives could lead to a lack of interoperability, duplicated work, and gaps in areas that are key for operators.

Telco networks require a higher level of control and assurance that general-purpose AI systems and agents do not typically provide. Agents need context, defined permissions, secure identity, clear policies, observability, and a way to validate their decisions before acting. Operators also need mechanisms to intervene, override, and understand what happens when an agent acts. These requirements become even more critical if agents start taking actions on live network infrastructure.

NGMN's review of the ecosystem suggests there is substantial activity across standards bodies and open-source communities, but the work is spread across many organizations and is developing at different speeds. Without greater alignment, operators will face competing approaches to agent communication, information models, context sharing, identity, and governance. This fragmentation could become a major constraint on adoption because operators cannot build an autonomous network around a collection of agents that cannot share context or interact with different suppliers. The industry needs to make the agents work together before they can coordinate the network.

That makes the supporting ecosystem almost as important as the agents themselves. Network equipment vendors have access to the orchestration and management platforms that agents will need to access. Hyperscalers bring AI infrastructure and agent platforms, and independent vendors can build systems designed to work across multiple vendor environments. The companies that can make these systems interoperable will become important players in the transition toward Level 4 autonomy.

NGMN's report creates an important opening for vendors that can connect existing network automation to Agentic AI. Network equipment vendors already have the controllers, orchestration platforms, and data that agents will need to access, while hyperscalers and independent AI vendors have strengths in agent platforms and AI infrastructure. An important example of this application is the Service Management and Orchestration (SMO) function in RAN networks. These companies must now focus on making their systems accessible by agents and demonstrating how agents can coordinate network functions across vendors. For vendors, this is likely to be more commercially important in the near term than adding more agents or capabilities.

Early deployments should be used to test how agents interact with existing controllers, what network context they actually need, and where human approval still needs to sit in the workflow. Interoperability will be important, as an agent that only works within one vendor's environment has limited value for an operator running a multivendor network. These deployments will also give operators data points for pushing vendors and standards organizations on the interfaces, governance, and trust mechanisms if systems are not up to standard.

However, by doing so, large infrastructure vendors may sacrifice a certain level of “stickiness” or the lock-in they have worked so hard to achieve. But the rate of AI development is relentless and vendors that opt to remain closed to their own domain are at risk of being left behind and replaced by more agile vendors. ABI Research expects that opening up for agentic network automation will quickly become a standard for all telco network infrastructure. This will require a shared effort across MNOs, network vendors, hyperscalers, standards organizations, and open-source communities, but network equipment vendors are best positioned because they control many of the existing interfaces and management systems that agents will need to reach. And most importantly, Agentic AI can provide the missing coordination capability needed for Level 4 autonomy, but the industry needs to build the environment around the agents first.