Huawei GTS teaches its ops agent to 'watch' networks troubleshoot, nearly clears dual-firewall test

Network troubleshooting is turning into a proving ground for AI agents, and Huawei's Global Technical Service (GTS) unit is betting on an agent that can 'watch' a network while it diagnoses faults.

According to qbitai, the approach nearly cleared the hard dual-firewall scenario, lifting task pass rate by 24.2% and cutting token cost by as much as 45%. Better accuracy and lower cost together is exactly the signal operators need before moving agents from pilots into production.

Troubleshooting is hard because the evidence is scattered. Topology, device configuration, logs and alarms live in different systems, so engineers have to compare views, form hypotheses and verify them one by one, and a single wrong step sends the investigation down the wrong path.

The phrase 'watching the network' is the interesting part. It suggests the agent is not just replaying a text command sequence but has to make sense of interfaces and live network state, which raises the bar for multimodal perception and for the stability of tool-calling chains.

Dual-firewall setups typically involve policy interplay across more than one device, where the root cause rarely sits at a single point and has to be found by repeatedly cross-checking multiple views. That work has long depended on experienced engineers and is among the hardest things to cover with conventional automation scripts.

The 45% token-cost reduction matters just as much. Network operations tasks are frequent and long-chained, so inference cost directly determines whether a deployment can scale; a lower cost per task is what lets agents take over more routine work.

What to watch next: whether the capability generalizes beyond a specific scenario to more complex cross-domain networks, and whether Huawei GTS packages it as a product for carriers and enterprise customers.