.jpg)
The history of computer networks left us an unforgettable lesson: the OSI model did not win the implementation battle, but it did win the language one. Today, as agentic artificial intelligence begins to weave its own communication infrastructure, we find ourselves in an analogous moment. Protocols like MCP, A2A, ACP, AG-UI, and ANP are already in production, allowing AI agents to access tools, collaborate with each other, and interact with humans. However, a shared reference model to order this ecosystem is missing. This article proposes a reflection, not a literal mapping, on how these protocols could be understood from a layered perspective, and what gaps remain to be resolved.
Instead of copying the structure of the OSI model, it is worth asking what network functions agents really need. Below, the physical infrastructure —GPUs, inference servers, AWS and Azure cloud services— constitutes the foundation on which everything rests. Here, Q2BSTUDIO offers AWS and Azure cloud services that allow scaling these workloads without worrying about the physical layer. At the next level, transport is already solved: A2A uses HTTP, AG-UI uses WebSockets. Nobody reinvents the wheel, and that is sensible.
Agent discovery and addressing is where the first real innovation appears. A2A uses 'Agent Cards' (JSON files at known URLs) that function as a lightweight DNS for agents. ANP goes further with decentralized identifiers (DIDs), allowing agents to authenticate without relying on a central registry. Task management —equivalent to the reliable transport layer— materializes in explicit states: submitted, working, completed, failed. It is like having a TCP for units of work, not for packets.
Where the parallelism with OSI breaks down is in the session layer. ANP introduces a 'meta-protocol' in which two agents negotiate in natural language how they will communicate. Routers never did this. It is a sign that agentic AI not only inherits old patterns but creates new ones. The presentation layer is reflected in the standardization of formats: MCP defines schemas for tools and resources, ACP uses MIME types. Both solve the same problem: that both parties agree on the form of the data before interpreting it.
At the top, the application layer hosts most of the discussions. A2A and ACP orchestrate machine-to-machine communication; AG-UI manages human interaction. Mixing them in the same bag obscures that they pursue different ends. For companies developing AI for businesses, understanding these differences is critical when designing custom applications that integrate agents.
However, there are three gaps that the OSI model never contemplated. Identity is the most evident: OSI assumed a trusted network. Agents need to prove who they are without shared history; ANP solves this with digital signatures and DIDs. Machine-to-machine payment is a problem that did not exist in 1984. Today, an agent may require compensation before responding. Some proposals reuse HTTP status code 402 (Payment Required) as a trigger for a settlement step. There is no equivalent in OSI. Governance is the thorniest: if an agent misbehaves, who is responsible? OSI always left responsibility with the human operator. With autonomous agents, that stops working.
These gaps are not minor. The convergence of agentic protocols will depend on how they are resolved. Meanwhile, organizations that want to get ahead need technology partners who understand both infrastructure and logical layers. Q2BSTUDIO, with experience in custom software, cybersecurity, business intelligence services, and Power BI, is prepared to help integrate AI agents into complex enterprise ecosystems. The future of agentic AI will not be written with a single standard, but with a pragmatic combination of protocols and good practices, just as the internet found its way beyond OSI and TCP/IP.
