AI agent libraries risk becoming telecom’s new OSS trap

Agent libraries are becoming the new OSS lock-in for telecom operators, and the AI model underneath is not the risk—the bus is.

Nokia’s Gemini-backed Agent Library, Samsung’s Agent Fabric inside CognitiV NOS, Amdocs aOS with its Cognitive Core, and Ericsson’s cApps on EIAP are each selling the same story: specialised agents that triage alarms, reason over KPIs, remediate, and push operators toward TM Forum L4.

Productivity claims cluster in the 50–80 percent range. Glass-box autonomy, intent, closed loops—the marketing is already hardening.

The risk is not which foundation model sits underneath; operators can swap models. However, they cannot cheaply swap the layer that registers agents, routes intent, holds the ontology, enforces guardrails, and decides which agent is allowed to touch the live network. That layer is the new OSS.

Choose the wrong bus and an operator has standardised on a vendor’s agent runtime the same way a generation of CSPs standardised on a vendor’s inventory, assurance, or order-management stack.

Four fabrics, one pattern

Nokia is shipping a starter pack of six Gemini agents (router, event triage, KPI selector, anomaly reasoner, action reasoner, and dashboard) all built on Google Cloud’s Agent Development Kit and landing as SaaS on Marketplace. The router is the orchestration plane, and more agents covering topology, services, and security follow. The library is prebuilt, domain-trained, and tied to Nokia Assurance Center and the Autonomous Networks Suite.

Samsung put an Agent Fabric at the centre of CognitiV NOS: a knowledge and decision layer where agents plan, evaluate, and act as a coordinated set rather than as separate point tools. It is already in operator networks. The pitch is end-to-end lifecycle automation toward L4/L5, with the fabric coordinating work across agents.

Amdocs calls aOS the first agentic operating system purpose-built for telecom. Cognitive Core supplies the telco agent library, ontology, and insights; aOS sits on top of any BSS/OSS and claims openness via REST, MCP, and A2A. “Open by design” is the slogan, but the library and the Cognitive Core remain Amdocs’ own.

Ericsson is extending EIAP from rApps into the core with cApps, and talking about an agent fabric as the control plane for discovery, governance, routing, and observability across OSS/BSS. Agentic rApps-as-a-service on AWS, Telco Agentic AI Studio, and Business Value Pathways position agents as first-class products on Ericsson’s automation platform.

Once a NOC, assurance, and service operations speak one vendor’s agent protocol, registry, and policy model, every new agent is cheaper to buy from that vendor than to integrate from outside.

The AI model is not the lock

Operators already know how to run multiple LLMs. What they do not have is a portable contract for what an agent is in a telecom production environment.

A useful agent is not a chat wrapper but a governed actor: typed tasks, SID-aligned vocabulary, topology and inventory context, policy bounds, audit, human-in-the-loop for high-stakes changes, and the ability to negotiate with other agents across RAN, transport, core, and BSS.

A2A-T, OpenAN runtimes, Trusted Data Gateways, and High-Value Scenario bindings try to make “any agent, any vendor, one fabric” real rather than a slide.

Until that contract is operational, each vendor’s library is a private dialect. Nokia’s router agent does not discover Samsung’s fabric agents. Amdocs’ Cognitive Core agents do not register cleanly on Ericsson’s EIAP. The CSP becomes the integration factory again (i.e. the OSS tax the industry tried to escape.)

Ericsson’s own agent-fabric writing is quite honest on this point: without a shared control plane, multi-vendor agents duplicate work, conflict on actions, and leave the operator stitching coordination by hand. That is the lock-in, described from the inside.

L4 marketing hardens the fabric early

L4 is the prize in every deck: intent in, closed loop out, humans on exceptions. Vendors need a packaged answer now because operators are buying “autonomous networks” as a programme, not as a protocol. Prebuilt agent libraries are the fastest path to a demo and a purchase order.

That speed is the trap. If procurement treats the agent library as a feature of the existing RAN/OSS vendor, the fabric hardens before interoperability does. In three years, the industry will have four incompatible agent buses—each with a thriving first-party catalogue, each claiming multi-vendor support via MCP or A2A, and all while the production path still runs through one registry, one ontology, and one governance agent.

We have seen this plenty of times before: inventory lock-in, assurance lock-in, and the “open API” that only fully works against the vendor’s own southbound. Agent libraries will be worse because they sit above the stack and accumulate decision context. Switching the bus means migrating institutional memory, not interfaces alone.

Interoperability here is the difference between an agent ecosystem and four walled gardens with L4 stickers.

Operators should treat the agent bus as a strategic control point and write it into RFPs now:

• Register once, run anywhere: An agent published to a vendor-neutral registry (capability, HVS binding, SID types, risk class) must be discoverable by any compliant fabric, with no private yellow pages.
• Typed telecom semantics, not prompt soup: A2A-T/TMF-aligned task, event, and negotiation types so “fix this cell’s uplink anomaly” means the same thing in Nokia, Samsung, Amdocs, and Ericsson runtimes.
• Portable policy and audit: Guardrails, approval gates, and decision logs must export. If the only explainability sits inside one vendor’s glass box, the operator does not own the autonomy.
• MCP and A2A as floor, not ceiling: Tool access and agent dialogue are necessary but not sufficient without telecom-grade identity, topology context, and closed-loop safety.
• First-party libraries as optional, not default: Buy the bus separately from the agents, and score vendors on how many third-party agents have run in production on their fabric rather than how many they shipped themselves.

The AI model war will keep moving, but the operator who standardises on one agent bus in 2026 will spend the 2030s paying to unwind it. Multi-vendor agent interoperability is how you stop the next OSS monopoly forming in plain sight.

See also: Samsung sets out agentic AI path to autonomous networks

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the IoT Tech Expo and Cyber Security & Cloud Expo. Click here for more information.

Telecoms is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.