Nokia brings Gemini agents to the network: automating without taking the engineer out of the loop

Manual network management has hit its ceiling. With traffic volumes that keep growing and increasingly tangled infrastructures, operations centers live buried in alarms, most of them noise. In that context, Nokia and Google Cloud have announced the integration of the Gemini models into the Nokia Assurance Center through six specialized agents: a router that orchestrates and interprets intent, an event-classification agent, a KPI selector, an anomaly reasoner that separates the real problem from the false alarm, an action reasoner that recommends fixes, and a dashboard agent that generates monitoring panels from natural language.

The architecture has a couple of design wins worth underlining. The first is practical: built on Google Cloud's Agent Development Kit and Gemini Enterprise Agent, it runs on the standard compute and storage the operator already has, without requiring complex managed services or new dedicated infrastructure. That lowers the barrier to entry and shortens deployment timelines, a decisive factor in a sector where every migration carries weight.

The second, and more relevant from the perspective of responsible AI, is what Nokia calls 'glass-box autonomy'. Instead of sidelining the engineer, the action reasoner agent works as an advisory layer: it proposes, and the human approves at the critical checkpoints before execution and logging. For low-risk scenarios previously authorized by internal policy, the same framework allows a closed loop without intervention. It is a configurable spectrum of autonomy, not an all-or-nothing switch, and it fits the growing regulatory demand for human oversight in sensitive decisions. It is, moreover, the honest way to introduce agents into critical infrastructure: earning trust in stages, not on faith.

There remains the chapter of the numbers. Nokia estimates reductions of between 50% and 80% in resolution times, along with fewer false alarms and escalations. They are attractive figures, but they should be read as the manufacturer's expectation: their validity will be measured in real deployments, with the heterogeneity and edge cases that only appear in production. If they are confirmed, even partially, the impact on network availability and operating costs would be considerable. The underlying signal is clear: the autonomous network stops being a slogan and becomes a roadmap with humans in the loop, which is exactly where they should be while trust is built.