NetCanvas: Huawei GTS Gives AI Agents an Interactive Visual Topology for Network Troubleshooting

The article opens with a quote from cognitive philosopher Andy Clark: true intelligence knows how to use external diagrams for "cognitive offloading." In complex IP carrier networks, troubleshooting requires a real-time mental map of dynamic topology — a skill only senior experts possess. Huawei GTS aimed to replicate this with LLM agents, but discovered a fatal flaw: agents understand logs yet suffer "topology amnesia" — they cannot retain the spatial layout of hundreds of nodes, leading to repeated probes and dead ends.

Problem: Topology Amnesia

In a real test, an agent identified the faulty firewall but spun for 300+ steps on that device, unable to locate its position in the multi-path topology. The root cause: pure text agents face a one-dimensional CLI stream, forcing them to reconstruct the network from fragmented logs at each step. As context grows, spatial relationships collapse.

Solution: NetCanvas — An Interactive Visual Topology

NetCanvas acts as an "external working memory" canvas, letting agents offload spatial reasoning. It operates via three mechanisms:

Dynamic Growth (Explore-as-you-grow): No static full-map upfront. Each CLI probe merges newly discovered devices, links, and path evidence into the live graph — like lifting fog of war.

On-Demand View & Reasoning Interaction: Agents switch global/local views, highlight focus nodes, and actively recolor links (green/red dashed) to mark connectivity hypotheses — enabling "look, mark, troubleshoot" loops.

Cognitive Alignment (Engineer-Centric Layout): Rendering topology in the layout conventions familiar to network engineers (domain layout) dramatically boosts success rates; the visual language must match human expert cognition.

Experimental Validation on CTBench

CTBench uses real maintenance cases requiring agents to enter unknown networks, execute commands dynamically, and locate faults from incomplete evidence.

Accuracy: Overall pass rate rises from 30.3% to 54.5% (63.6% with physical-link priors). On dual-firewall and ECMP sub-tasks — heavily dependent on spatial reasoning — performance leaps from ~10% to ~90%.

Stability: Zero regressions across the benchmark (16 wins, 0 losses, 50 ties).

Efficiency: Average probe steps drop from 159 to 113. Per-solved-task token cost falls 26–45%, proving the visual canvas reduces costly trial-and-error.

Broader Implications

The peripheral system (Harness) is the agent's extended mind. For topology-rich domains, stuffing massive logs into tokens guarantees cognitive overload. Professional agents need not only strong reasoning but an evolvable "extended working memory" that mirrors the expert's spatial cognition.

Huawei GTS's path: engineer the seasoned engineer's spatial cognition and work habits into the agent's perception-action loop. The project and benchmark are open-sourced:

NetCanvas: https://github.com/caimanjing/netcanvas CTBench dataset:

https://huggingface.co/datasets/netop/CTBench