Toward an OSI-inspired Stack for Agentic AI: From Fragmented Frameworks to Open Standards

Abstract

Agentic Artificial Intelligence (Agentic AI) is emerging as a paradigm where autonomous agents, often powered by Large Language Models (LLMs), collaborate with tools, data sources, and each other to accomplish complex objectives. Yet the promise of these systems faces a fundamental obstacle: interoperability. Current frameworks and protocols remain fragmented, forcing costly ad hoc integrations and preventing the emergence of scalable, multi-agent ecosystems. In this paper we argue that interoperability is not a secondary technical concern but the central challenge that will determine whether Agentic AI matures into robust digital infrastructure. To analyze this problem, we introduce an interoperability stack inspired by the OSI model in computer networking. The layered perspective clarifies responsibilities across seven levels, from physical infrastructure and runtime execution to agent-to-agent communication, coordination workflows, and governance and trust. Mapping today’s initiatives onto this framework highlights stark asymmetries: lower layers are relatively mature, middle layers remain underdeveloped, and upper layers —especially governance— are almost entirely absent. Building on this diagnosis, we outline a research and standardization agenda. Priorities include formal specifications for agent-to-agent communication, interoperable workflow schemas, benchmarks for measuring interoperability, and governance-by-protocol mechanisms that embed accountability and compliance into communication itself. We argue that, as with TCP/IP and the Internet, layered and open standards are the only path to prevent fragmentation and unlock the transformative potential of Agentic AI. Without them, the field risks repeating the stagnation of pre-Internet networking; with them, it may enable ecosystems of autonomous agents as consequential as the Internet itself.

Keywords