A Living Network Map Helps Agents Debug Complex Topologies

Introduction

Network troubleshooting agents often perform well when asked to interpret an isolated command or log fragment. The difficulty increases sharply in a live topology containing many devices, multiple routes, firewalls, and equal-cost multipath routing. In such settings, the agent must not only understand each observation; it must also remember where that observation belongs in the network.

Huawei GTS describes this failure mode as “topology amnesia.” A text-only agent may identify the relevant firewall or protocol symptom, yet lose track of the path on which the device sits. As the investigation grows, more local evidence can paradoxically lead to a weaker global picture. NetCanvas is designed to address that gap with an interactive topology serving as external working memory.

How the approach works

NetCanvas is not simply a static network diagram placed in the model’s context. It is integrated into the investigation loop.

• Incremental growth: Newly discovered devices, links, and path evidence are merged into the current graph after each CLI investigation.
• On-demand views: The agent can move between a global view and a focused local view, highlighting the node or route currently under examination.
• Visible hypotheses: Link colors and line styles can be used to record connectivity states or reasoning assumptions, keeping intermediate conclusions outside the text context.
• Domain-aligned layout: The team reports that arranging the graph in a way familiar to network engineers also matters. A technically correct but cognitively awkward diagram may be less useful to the agent.

This division of labor is important: the language model chooses what to inspect and how to interpret evidence, while the canvas preserves the evolving spatial structure.

Reported benchmark results

According to the supplied material, tests on the CTBench communication-operations benchmark raised overall task completion from 30.3% to 54.5%. With physical-link priors, the reported result reached 63.6%. The largest gains appeared in tasks involving two firewalls and ECMP, where completion reportedly rose from roughly 10% to roughly 90%.

The team also reports no per-task regression in its comparison: 16 tasks were newly solved and 50 remained unchanged. Average exploration steps fell from 159 to 113. When token trial costs were calculated per successfully solved task, the reported reduction ranged from 26% to 45%.

These figures come from the project’s disclosed benchmark evaluation. They indicate the potential value of interactive topology, but do not by themselves establish performance across vendors, larger production networks, or unseen operational conditions.

Why it matters

The broader lesson is not merely that agents should “see” diagrams. Expert network engineers continuously maintain a spatial model of traffic flow, branching points, and device position. Turning that model into a persistent, editable system state can reduce the burden placed on the model’s context window.

This suggests a wider design principle for professional agents: capability depends not only on the base model, but also on the harness around it—tools, state representations, and interfaces. Similar external working memories may be useful in databases, industrial procedures, and other domains with structured relationships. NetCanvas offers a concrete direction, while its long-term value will depend on topology quality, tool permissions, and coverage of real failure cases.

Project: caimanjing/netcanvas

Benchmark: CTBench

Source: QbitAI