Six specialized Nokia agents will roll out through 2027
In sum – what we know:
- A six-agent system – Nokia built a router, event triage, KPI selector, anomaly reasoner, remediation, and dashboard-generation agent, all coordinated by the router as orchestration layer.
- Glass box autonomy – Agents act on their own for low-risk routine work but require human sign-off for high-stakes changes to core networks.
- Rolling rollout – The router and triage agents are already live, with the full SaaS package landing on Google Cloud Marketplace in September 2026 and more agents shipping through 2027.
Google Cloud and Nokia have expanded their partnership to push Gemini-powered AI agents directly into Nokia’s network operations software, with the integration centered on the Nokia Assurance Center autonomous network suite. The deal embeds Google Cloud’s multimodal Gemini foundation models into the platform telecom operators use for network monitoring, service assurance, and day-to-day operations.
The two companies have been working together for a while, with Nokia launching three of its Network as Code (NaC) APIs on the Google Cloud Marketplace in July 2025, and the companies showcasing deeper agentic AI integration at MWC Barcelona. The Assurance Center announcement reads as a concrete implementation of that broader vision, focused specifically on operations and assurance rather than developer-facing APIs.
Technically, the agents run on standard Google Cloud infrastructure using Kubernetes and Google Cloud Storage, built with Google Cloud’s Agent Development Kit (ADK) on the Gemini Enterprise Agent Platform. That’s a notable point for Google in a telco cloud market where it trails Microsoft Azure and AWS for infrastructure workloads.
Six specialized AI agents
Nokia has built six agents, each with a defined role. The router agent sits at the center as the orchestration layer, interpreting natural language commands from operators, managing communication between the other agents, and enforcing operational guardrails so behavior stays inside policy boundaries. It’s what lets Nokia describe this as a coordinated ecosystem rather than a bag of disconnected tools.
The event triage agent ingests network alarms and events, weighs them against historical patterns and topology, and infers probable root causes. Its main job is to filter noise and surface genuine issues — a pain point in NOCs where operators can drown in thousands of alerts after something like a fiber cut or a botched software upgrade. The KPI selector agent tackles a different kind of complexity, offering domain-expert interpretation of fragmented performance metrics that vary across network layers and vendors, and helping operators pick the most relevant ones for a given problem. The anomaly reasoner complements triage by using historical context to separate genuine deviations from benign variations, which matters for cutting down on unnecessary escalations and the operator fatigue that comes with them.
Rounding out the set, the remediation agent generates fixes such as traffic re-routing or configuration changes and can execute low-risk actions automatically, while higher-risk ones require sign-off. The dashboard-generation agent assembles dynamic visualizations from telemetry, KPIs, and historical data, which is also where Gemini’s multimodal capabilities show up in a telecom context.
The pitch against Ericsson and Cisco, both of which have their own AI/ML-based operations tooling, is the breadth and specialization of this multi-agent setup rather than a single monolithic assistant. Whether a coordinated ecosystem genuinely beats a tightly integrated proprietary stack is the kind of thing that only shows up in production. And, of course, that’s not to mention the fact that presumably others could scope out multi-agent systems too.
Rolling rollout timeline
Two of the six agents — the router and event triage — are already fully functional and running inside Nokia’s environment, forming the backbone that the others plug into. The full platform is slated to launch as a software-as-a-service package on the Google Cloud Marketplace in September 2026, with operators able to deploy an initial starter pack of certified agents at launch.
Rather than holding everything back until all six agents and their supporting capabilities are complete, Nokia plans to ship the more complex pieces via rolling software updates through late 2026 and into 2027. Those updates will extend agent-based automation across Nokia’s wider portfolio, including Unified Inventory for asset and resource management and the Data Suite for integration and analytics. That’s a multi-year roadmap, not a one-off product drop.
The transparency of a hyperscaler-backed SaaS timetable is part of the competitive story here, particularly against Huawei’s AI-empowered network management tools deployed across the markets where it operates. A published rollout schedule and a marketplace distribution model are easier for Western operators to evaluate and procure than tools tied to a different geopolitical footprint.
“Glass box autonomy”
Nokia keeps returning to a phrase it calls “glass box autonomy,” set up in deliberate contrast to “black box” AI. The idea is that human engineers retain final approval over critical decisions and actions, while the system handles fully automated closed-loop responses only for lower-risk scenarios like routine adjustments and non-critical optimizations. In other words, the AI can act on its own where the stakes are low, and asks permission where they’re high.
The framing is clearly built to address the things that make operators nervous about handing operational control to AI. There’s the operational risk of an agent making an unintended change to a core network, the regulatory pressure for human oversight and auditability — sharper in Europe than most places — and the basic need for explainability, meaning an operator can understand why an agent recommended or executed a given action. These are reasonable anxieties, and pre-built explainability is genuinely useful for operators who’d otherwise have to construct their own data science and governance scaffolding around in-house tools.
The headline number attached to all of this is a claimed 50% to 80% reduction in network problem-solving times, with issues that used to take hours supposedly resolved in minutes. We’ll have to wait and see how that pans out once all of the agents in the system are fully deployed.
Table of Contents
