From Automated to Autonomous: A new blueprint for Level 4/5 networks

Lead: a shift from scripts to agents

A short letter on arXiv proposes a significant conceptual shift: move networking from fixed automation to "agent-native" intelligence. The Hierarchical Agent-native Network Architecture (HANA) argues that reaching Level 4/5 Autonomous Networks (AN) — networks that can plan, reason and recover without human intervention — requires replacing brittle scripts with cognitive multi‑agent systems. The paper, arXiv:2605.20608, lays out a hierarchical multi‑agent reference architecture intended to give networks the agency to handle off‑nominal conditions that current operations cannot.

What HANA proposes

HANA describes layered agents with distinct roles — high‑level strategic coordinators, mid‑level orchestration agents, and low‑level execution agents — that cooperate to achieve policy goals, maintain safety envelopes and learn from anomalies. The architecture emphasizes agent-native primitives rather than retrofitting autonomy onto legacy automation stacks. It is presented as a reference design: a roadmap for researchers, vendors and operators to build networks that can adapt, reason about intent and coordinate across domains (edge, cloud, 5G/6G slices).

Why this matters — and the broader context

Why should Western readers care? Autonomous networks could reshape how telecoms, cloud providers and critical infrastructure are managed, reducing human toil while raising new questions about control, transparency and provenance. There are geopolitical stakes too: it has been reported that regulators and governments are increasingly scrutinizing autonomous systems in critical infrastructure, and ongoing export controls and sanctions on advanced AI accelerators will affect who can deploy these systems at scale. Who designs the agents, who audits them, and under what legal regimes they operate will matter as much as the technical design.

Next steps and caveats

HANA is a research proposal, not a deployed system. Real‑world validation, standards work and safety‑first governance will be required before networks can claim Level 4/5 autonomy. Trials with operators, open specifications and independent audits will determine whether agent‑native designs deliver resilience without introducing unacceptable new risks. Reportedly, the conversation is only just starting — and the architecture lays down a clear place to begin.