Distributed computing keeps cycling through the same pattern: competing standards emerge, then one wins through simplicity. REST beat CORBA, DCOM, and Java RMI because it worked with HTTP. MQTT and WebSockets became the go-to protocols for messaging and real-time data after XMPP and IRC fragmented the space.
The article draws a parallel to recent AI infrastructure developments. Model Context Protocol (MCP) solved how AI agents call tools by establishing a unified interface. A2A (agent-to-agent) protocols solved coordination between multiple AI systems. But transport, the article suggests, remains unsolved.
Transport refers to the underlying mechanism for moving data between systems. In the REST era, HTTP became the obvious choice because it was already ubiquitous. Today's AI infrastructure lacks an equivalent consensus. Developers use whatever works for their specific case, whether that's REST APIs, gRPC, WebSockets, or message queues. This fragmentation creates friction when integrating AI systems at scale.
The parallel holds weight. Early distributed computing needed a standard protocol to let different systems talk. That problem solved itself once REST's simplicity proved its worth over competitors. AI tooling now faces a similar inflection point. MCP standardized what tools AI models can call. A2A frameworks began standardizing how agents coordinate. But the pipes carrying all this data lack equivalent standardization.
What would win? Likely something that mirrors REST's advantage: simplicity married to an existing standard. HTTP remains ubiquitous. WebSockets handle real-time cases efficiently. gRPC offers performance where it matters. The question isn't whether any of these will dominate, but whether the market will consciously converge on one rather than letting fragmentation persist.
The stakes differ from CORBA versus RMI. AI infrastructure touches fewer organizations than enterprise integration did. But consolidation will