Everyone else is also under huge pressure to get agentic AI to market, whether that is client-facing or internal to the organisation. And they're all worried about what happens the next day too – because they also know the protocols aren't up to the job yet, and disaster seems all but inevitable.
That's as good as the news gets: it's not just you.
On Tuesday, an obscure breakaway session at Kong's API Summit Live in New York drew a dangerously overcapacity crowd with the title "MCP vs OpenAPI vs A2A vs ?: Preparing for the Agentic World".
It is not that the attendees to an API conference tagged as "for the agentic era" don't understand the differences between the Model Context Protocol introduced by Anthropic and Google's Agent2Agent Protocol, presenter Greg Peranich told The Stack.
It is that they, from engineers to top-level management, "don't know where to start".
A2A is six months old, and MCP turns one towards the end of next month. But for those who want to push products to production, they're mostly trouble.
It is, said Kong staff solutions engineer Peranich, "rather difficult" to work with MCP when the spec changes almost weekly. A2A, meanwhile, is far more complex than the traditional APIs developers know and love, while offering no standard for telemetry, logging, and audit.
But sticking to OpenAPI won't cut it.
APIs: for the humans
Some organisations have the high-level idea of giving agents access to whatever tools or data they need via APIs, and manage on that level.
And indeed, said Peranich, OpenAPI has a mature ecosystem and makes life "super easy for devs". But it is human-centric, with rigid schemas, a distinct lack of intent-based reasoning, and poor semantic context.
MCP offers rich context, it is lightweight – and it is agent-native. A2A, meanwhile, serves the needs of agents that need to negotiate, delegate, and coordinate.
But both lack standards for telemetry, logging, and audit. Both have issues with authentication. Both can struggle when there is a human in the loop.
So how do you build secure and responsible systems using such early-stage protocols? Nobody else knows either.
Then there are open questions such as lifecycle management, said Peranich, plus semantic metadata and ontologies.
Gateway to a polyglot world
The big question, of course, is whether agentic AI protocols converge towards a single standard, or whether the future is irredeemably polyglot.
Peranich leans towards polyglot forever, and Kong's product evolution concurs. On Tuesday it announced a set of MCP capabilities for its flagship products to surveil and govern MCP servers centrally.
If you can abstract and keep an eye on a zoo of agents, you can exert some control, said Peranich.
But if you're looking for a protocol that will make autonomous interactions between agents as easy as it is to make dumb systems talk to one another, best of luck to you.
The link has been copied!
