A new preprint sketches a vision of heterogeneous AI agents discovering, negotiating, and collaborating across platforms. The real story is the catalog of problems standing in the way.
Imagine a world where AI agents from different companies, platforms, and research labs could discover each other on a network, negotiate responsibilities, exchange context, and execute shared workflows the way browsers and servers do on the web. That is the future a new preprint posted to arXiv lays out under the banner of an "Internet of Agentic AI." The paper is not a blueprint for that world. It is closer to a diagnostic of how far the field is from it.
The paper's single author, Quanyan Zhu of NYU Tandon School of Engineering, describes an open ecosystem where heterogeneous agents, ranging from cloud services to edge devices to cyber-physical systems, would participate in coordinated activity without a central operator dictating their interactions. The proposed vision borrows from single-agent agentic AI, multi-agent systems, distributed computing, communication networks, game theory, and security engineering, and walks through agent deployment models, workflow lifecycles, communication protocols, interoperability layers, resource management, and trust architectures. Two case studies, in adaptive manufacturing and distributed operational coordination, illustrate what such a system might look like in practice. The paper presents these as illustrations of the framework rather than as evidence of a working system.
Read in isolation, the framework sounds like a near-term engineering target. It is not. The paper's own value is in the list of research challenges it surfaces, which read less like a research checklist and more like a confession of what the field has not yet solved. Among them: semantic interoperability (how do agents built on different models, by different vendors, agree on what a request means?), secure identity (who is this agent, and who is accountable for its actions?), incentive-compatible coordination (why would competing services cooperate?), resource-aware orchestration (how do you schedule work across agents with wildly different compute, latency, and cost profiles?), and governance of emergent behavior (what happens when dozens of agents acting rationally produce a system outcome no one intended?). Zhu frames these as open research problems rather than solved engineering tasks.
Each of these is a known hard problem dressed in the paper's vocabulary. Semantic interoperability is the same hurdle that has tripped up every previous generation of inter-vendor data exchange, from EDI to SOAP to modern API gateways. Zhu is not pretending to solve it. The argument is that it has to be solved before a multi-agent internet can function, and that the agent era raises the stakes because the "data" being exchanged is intent, context, and partial plans rather than typed records.
The paper also lands in a landscape that is already moving. Adjacent protocol efforts, including Anthropic's Model Context Protocol (MCP) and Google's Agent2Agent (A2A) initiative, are early attempts to standardize how agents expose capabilities, share context, and hand off tasks. MCP, an open-source standard Anthropic publishes, lets AI applications connect to external data sources, tools, and workflows. A2A, an open protocol Google's A2A project publishes, is designed to enable communication between opaque agentic applications built on different frameworks by different companies. Neither is a full "Internet of Agentic AI" in the sense the preprint describes, but they occupy parts of the same design space — a proximity the paper does not itself cite or claim, but which frames its open research questions in a landscape where some of those same problems are already being treated as engineering problems by industry. A serious roadmap has to acknowledge that the gap between "research agenda" and "deployed protocol" is narrower than the paper's framing implies, even if it is still wide.
What would it actually take to close that gap? On the protocol side, the field would need shared semantics for tasks and capabilities, a way to authenticate and authorize agents across organizational boundaries, and resource-discovery mechanisms that work at internet scale rather than within a single vendor's platform. On the engineering side, agents would need reliable ways to fail safely when a peer misbehaves, which means standardized error reporting, sandboxing, and rollback. On the governance side, someone would have to answer the accountability question: when an autonomous agent takes a consequential action on a user's behalf, who is the responsible party — the user, the agent's operator, the platform that connected them, or the model provider? No current regulatory or contractual framework answers this cleanly, and the preprint does not pretend otherwise.
The paper is best read as a research agenda, not a forecast. Its contribution is the catalog of open questions, not a claim that the questions are answered. For anyone building, buying, or regulating AI systems, the practical takeaway is that the most consequential decisions in the agent era will be made at the protocol and governance layers, not inside any single model. The "Internet of Agentic AI" is, for now, a destination. The preprint is a map of the road that has not yet been built.