Huawei fixes its network agent's 'topology amnesia' with a live map
Two stories out of China's phone-and-carrier world this morning, and both make the same bet: the next agent gains come from the harness around the model, not the model itself.
Huawei's GTS AI team says it cured its troubleshooting agent of "Topology Amnesia" by giving it a map that updates as it probes. The diagnosis is sharp: a text-only agent can parse any single CLI output fine, but deep into a multi-device, multi-path network with firewalls and ECMP in the mix, the local observations pile up while the global spatial picture collapses — one test run locked onto the right firewall, then burned 300-plus steps spinning on that one box. The fix, NetCanvas, is a runtime layer that merges every probe into a live graph state and renders the relevant slice back to the model, which can zoom, focus, and draw its own hypotheses onto the map as red/green dashed hop markings — cognitive offloading, done literally. On CTBench, a public benchmark of fault-localization tasks built from real operations cases (first 66 problems), the same model and harness, with only the working-memory format changed, jumped from 30.3% pass to 54.5%, or 63.6% when physical wiring is given as prior; parallel-path cases went from 10% to 90%, average tool calls fell from 159 to 113, and QbitAI reports token costs down as much as 45%. The team is honest about the boundary: all 66 problems ran on one enterprise-campus-plus-star-topology network, and whether the domain-specific layout trick holds in spine-leaf data centers or dense multi-tenant clouds is untested — the remaining failures were extraction and multi-hop reasoning, not drawing.
vivo launched four Blue Heart models and a system-level "Blue Heart Harness" at its developer conference today. Blue LM-RealTime handles end-to-end speech understanding in noise, accents, and mixed Chinese-English; Blue LM-Nano runs on-device, accumulating personal memory from screen content, files, and usage habits; Blue LM-Flash and Blue LM-Pro sit in the cloud and act as the phone's task controller, calling across apps and devices to actually complete requests. It lands two days after we covered Honor's system-level agent harness — Honor ships the first phone with a system-level agent harness — and Chinese OEMs are now converging on the same playbook: own the OS layer where the agent lives, rather than renting it to a foreign assistant.
What to watch: NetCanvas is published with a paper, repo, and the open CTBench dataset — worth stealing the evaluation design for any agentic ops benchmark you run.
Is your agent's real bottleneck the model — or the map you give it? Tell us in the comments.
Sources: QbitAI · NetCanvas project page (Huawei GTS team) · CTBench paper (arXiv) · NetCanvas (GitHub) · AIbase · Sina Finance
