What Is the Agent Internet? A Network Perspective on AI Agent Interconnection
From an Internet of Humans to an Internet of Agents
Today’s Internet was designed for people: browsers, apps, account systems — every protocol assumes a human at the other end. But as LLM-powered AI Agents proliferate, a new kind of “resident” has appeared on the network: always online, issuing requests at machine speed, needing to collaborate with each other, and even acting on behalf of humans to invoke other agents.
When billions of agents start communicating, collaborating and transacting with one another, a question emerges: do they need an Internet of their own? That is the starting point of the “Agent Internet” (or Agent Web) concept.
What Problems Must It Solve?
From a network architecture perspective, at least four:
-
Identity and addressing. Humans and web pages have URLs — what about agents? An agent may be deployed in the cloud, at the edge, or even migrate across hosts. Giving agents a stable, verifiable, routable network identity is the first-order problem. The vast address space and extension headers of IPv6 provide a natural proving ground.
-
Discovery and collaboration. How does an agent find the right partner? Today this is done through proprietary vendor directories; an open discovery protocol is needed — something like DNS for capabilities.
-
Trust and security. Agents act on behalf of humans. How are permissions granted, behaviors audited, and malicious agents isolated? Traditional ACLs and TLS are insufficient; agent-oriented identity, capability attestation and policy control are required.
-
Quality of service and accounting. Agent traffic has differentiated latency and bandwidth needs (e.g., bursty context transfer for inference). How can carrier networks become agent-aware and offer differentiated transport?
Standardization Progress
Standards bodies have moved. In the IETF, multiple Internet-Drafts address agent gateway inventory and control, IPv6 agent-aware networking frameworks, and inter-agent communication protocols. W3C and 3GPP are following from the web-protocol and mobile-network angles respectively. On the industry side, Anthropic’s MCP (Model Context Protocol) and Google’s A2A (Agent2Agent) are emerging as de-facto standard contenders.
A notable divide is between the end-to-end camp (agents negotiate and communicate directly, P2P-style) and the gateway camp (communication mediated by network-resident agent gateways providing unified identity, authentication and control) — the latter clearly aligns with the operational tradition of telecom carriers.
Summary
The Agent Internet is still at the “blind men and the elephant” stage: AI vendors, Internet companies and telecom operators each have their own vision. What is certain is that it will not appear out of thin air, but evolve on top of the existing IP network — just as the Web did over TCP/IP. For network researchers, this is an architecture-design window that comes once in decades.
This article reflects personal views only. Latest standards developments can be found in the Library.