The telecom industry just declared agent infrastructure its central thesis. At MWC 2026, Huawei launched agent registration at the network layer, NVIDIA assembled a twelve-company coalition for AI-native 6G, and ZTE built an industrial agent studio. The physical infrastructure for billions of AI agents is being constructed right now. The network will verify that an agent exists. It will not verify that a human approved the agent's actions.
MWC Barcelona opened this week with a thesis that would have been unrecognizable two years ago. The telecom industry's largest annual gathering — over a hundred thousand attendees, two thousand exhibitors — organized its headline programming around a single question: how do you build a network for AI agents?
The answer came from three directions in a single day.
What Was Announced
Huawei launched what it calls the Agentic Core — a triple-engine architecture that embeds agent registration, discovery, digital identity, and agent-to-agent session management directly into the telecom network layer. The same layer where phone numbers, SIM cards, and IP addresses live. The company is also open-sourcing A2A-T, an agent-to-agent communication protocol designed specifically for telecom infrastructure. The technical specifications are concrete: a hundred megabits per second of throughput and twenty milliseconds of latency to support AI-powered robots and autonomous vehicles. The system anticipates a tenfold increase in connected entities as agents become standard capabilities in next-generation devices. The network shifts from predefined rules to what Huawei describes as intent-driven operation — network AI agents dynamically matching resources to demand rather than following static configurations.
NVIDIA assembled a coalition. Twelve organizations — BT Group, Deutsche Telekom, T-Mobile, SoftBank, SK Telecom, Nokia, Cisco, Ericsson, Booz Allen, MITRE, and two others — committed to building sixth-generation wireless networks on open, secure, AI-native platforms. Jensen Huang framed the ambition simply: transform the world's telecom networks into AI infrastructure everywhere. The timeline is concrete: 6G trials in 2028, commercial deployment by 2030. T-Mobile and Nokia have already tested GPU-accelerated AI workloads on NVIDIA hardware at T-Mobile's AI-RAN Innovation Centre in Seattle, running video streaming, generative AI queries, and AI-based video captioning concurrently on a single server. NVIDIA also released Nemotron, an open-source large language model designed specifically for telecom network operations — the first foundation model purpose-built for agents that manage wireless infrastructure.
ZTE launched Co-Sight AI Agent Studio, an industrial-grade agent platform that ranked first in both the GAIA and HLE evaluation benchmarks. Its AIR MAX architecture is a ten-block AI stack supporting AI-native infrastructure, Level 4 autonomous network operations, and a revenue monetization engine. ZTE also introduced Co-Claw, a full-stack enterprise AI agent deployed across office operations, R&D, and network management. The operating paradigm ZTE describes: AI serves AI.
The Agentic AI Summit — a dedicated four-hour track on March 2 — organized presentations from China Mobile, AT&T, Orange, NVIDIA, Cohere, MIT, Palo Alto Networks, and AWS across four pillars: network autonomy, value creation, industry challenges, and trust. MIT will present its NANDA framework for decentralized agent governance. One session is titled 'Post-Human Network Architecture.'
What the Network Knows
Telecom networks have always been identity systems. A phone number is not just a routing address — it is a registered identity on the network. A SIM card is not just an access mechanism — it is a cryptographic binding between a physical device and a network-level credential. When a call is placed, the network knows who is calling. Not because the caller announced it — because the infrastructure verified it at registration.
Huawei's Agentic Core is building the same architecture for AI agents. Agent registration means the network knows this agent exists. Agent discovery means other agents and services can find it through the network. Digital identity means the agent carries a network-level credential. A2A session management means agents can establish verified, low-latency communication channels through the infrastructure itself.
All of this operates at Layer 3 — the network layer. The same layer where IP routing decisions are made, where physical infrastructure meets logical addressing, where the fundamental plumbing of connectivity is built. When an agent registers on a network running Huawei's Agentic Core, the network can verify three things: this agent is real, it is registered, and it has permission to consume network resources.
The network cannot verify a fourth thing: whether a human approved what the agent is about to do.
Two Layers, Two Questions
The distinction is structural, not semantic. Network-layer identity answers: is this entity authorized to be on the network? Application-layer authorization answers: is this entity authorized to perform this specific action?
The historical parallel is exact. SIM cards and passwords have coexisted for three decades. The SIM card verifies: this device is authorized to connect to this carrier's network. The password verifies: this person is authorized to access this bank account. Different layers, different questions, different trust models. Neither subsumes the other. The security of the entire mobile ecosystem depends on both layers operating independently.
Phone numbers provide network identity. Passwords provide application identity. They serve fundamentally different purposes, and removing either one does not strengthen the other.
Agent identity is splitting along the same architectural line. Huawei, NVIDIA, and ZTE are building the network-layer half. Agent registration is the new SIM card — proof that the entity is legitimate and permitted on the network. Agent discovery is the new DNS — a mechanism for finding registered entities by capability or name. A2A session management is the new circuit establishment — verified channels between known parties.
The application-layer half — verifying that a specific human approved a specific action at the specific moment of execution, not at credential issuance, not at account setup, but now — is not being built at MWC. No telecom company announced it. No coalition committed to it. The Agentic AI Summit's trust panel covers security, identity, and accountability. Identity at the network layer. Not authorization at the action layer.
Why the Gap Is Architectural
The gap is not an oversight. It is a consequence of what networks are designed to do.
Networks answer questions about connectivity and resource access. Is this device authenticated? Does it have bandwidth allocation? Is the traffic routed correctly? These are binary questions with deterministic answers, resolved at machine speed, verified by cryptographic protocols operating below the application layer.
Authorization at the action layer is a different category of question entirely. Did this specific person, identified by something they are rather than something they have, approve this specific action? Was the action they approved the same action that was executed? Can that approval be verified after the fact? These questions require identity binding that networks are not architected to provide — not because the engineering is hard, but because the abstraction boundary is wrong. A 5G tower can verify that an agent's traffic is authentic. It cannot verify that the email the agent sent was something its owner actually wanted sent.
This is the same structural reason that SSL certificates do not prevent phishing. The certificate verifies the connection is encrypted and the server's identity is authentic. It does not verify that the website is trustworthy. Different layer, different question. The certificate operates at the transport layer. Trust operates above it.
The Timeline
The scale is worth absorbing. NVIDIA's coalition includes operators serving billions of subscribers across six continents. Huawei's network equipment runs in over a hundred and seventy countries. The 6G standard will be designed from inception with AI agent support — not retrofitted, not patched in, but native. When this infrastructure goes live commercially around 2030, the number of registered agents on global telecom networks will likely exceed the number of human subscribers.
The infrastructure is being built now. The registration protocols are being standardized now. The agent-to-agent communication channels are being defined now. NVIDIA's open-source Nemotron model means network operations agents can be deployed by any carrier. Huawei's open-source A2A-T protocol means agent communication standards will emerge from telecom, not just from software.
The assumption embedded in all of it — visible in the architecture, implicit in the protocol design, absent from the summit agenda — is that someone, somewhere, at some other layer of the stack, will solve the authorization problem before the agents arrive.
The agents, based on the MWC announcements, are arriving in 2028.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)