Huawei GTS's NetCanvas Lets AI Agents 'See' Networks While Troubleshooting

IP transport networks carry digital services from card payments to online hospital registrations. When they fail, clues are often buried in massive alarms or silent configuration data. The first barrier is not how to repair the fault but how to determine what the network looks like at that moment. Real production networks are live and dynamic, and every failure creates a different topology. Only a small number of senior experts can reconstruct the changing network in their heads and identify the root cause.

The industry expected LLM agents to change that. With strong knowledge understanding, logical reasoning, and analytical decision-making, agents were seen as a natural fit for network operations. But after deploying agents in field scenarios, Huawei GTS's algorithm team found a counterintuitive problem: an agent could read large volumes of text logs yet could not remember the network's shape. The team calls this "Topology Amnesia." A text-only agent trying to hold hundreds of dynamic nodes in memory is bound to suffer memory collapse.

In one real troubleshooting test, the agent had already identified the faulty firewall, but it tried dozens of query methods on that device and spun through more than 300 steps without finding the root cause. It knew the firewall was relevant, but not where the device sat in the overall topology or which hop of which path it occupied. Senior network engineers keep a live map in mind: where traffic enters, where it branches, and where the firewall blocks it. A text-only agent faces a one-dimensional scrolling command-line interface, so every step requires rebuilding the network from fragmented logs. As context grows, those spatial relationships collapse.

NetCanvas was designed to solve that. Instead of writing expert experience into a prompt, the team wrote the expert's working method into the agent system. The model handles reasoning and decisions, while NetCanvas handles memory and drawing. It is not a static screenshot of a network topology but an interactive visual topology. The team describes three core functions.

First, the map grows as the agent probes. NetCanvas rejects the idea of providing a full topology at the start. Each time the agent completes a command-line probe, newly discovered devices, links, and path evidence are merged into the current graph state. The topology grows with the troubleshooting process, like fog being cleared step by step.

Second, the agent can view and reason with the map on demand. It can switch between global and local views, highlight a focus node, and change link colors, such as using green or red dashed lines to mark connectivity or record reasoning hypotheses. The model no longer has to memorize everything internally; it can look at the map, mark it, and troubleshoot at the same time. Third, the map follows network engineers' cognitive habits. In tests, replacing a machine-generated topology with a domain layout familiar to human network engineers produced a sharp increase in the agent's success rate. Drawing the map in the right way helped the AI take the right path.

The team tested NetCanvas on CTBench, a public communications operations benchmark. Its tasks come from real operations cases and require an agent to enter an unknown network, execute commands dynamically, and locate faults from incomplete, fragmented evidence. On the full dataset, NetCanvas raised the agent's overall pass rate from 30.3 percent to 54.5 percent; with a physical wiring prior, the rate reached 63.6 percent. On complex subtasks such as dual-firewall and ECMP troubleshooting, which depend heavily on spatial relationships, performance jumped from about 10 percent to about 90 percent. Compared with a text-only baseline question by question, NetCanvas recorded 16 wins, 0 losses, and 50 ties, meaning the extra successes did not come at the expense of other tasks. The agent's average number of exploration steps fell from 159 to 113. Calculated per correctly solved task, token trial-and-error costs dropped 26 percent to 45 percent.

The team said the results point to a broader lesson for professional agents: internal model reasoning matters, but the surrounding system, or harness, determines how well the agent can act in a complex real-world environment. For IP networks woven from topology, paths, and protocols, forcing massive logs into a model's token context causes cognitive overload. The agent needs an evolving external working memory. Huawei GTS's approach engineers the spatial cognition and working methods of senior engineers into system capabilities that an agent can perceive and interact with. The project is available at github.com/caimanjing/netcanvas, and the CTBench dataset is at huggingface.co/datasets/netop/CTBench.