Understanding AI Communication Protocols (and Why None Are Perfect Yet)
A goal without a path is just a wish. The same can be said for AI agents: without a way to talk to each other, they’re like brilliant minds trapped in separate rooms.
Aug 13, 2025
AI
5 min
The goal of an AI communication protocol is pretty ambitious:
Agents should be able to discover each other online.
Authenticate each other.
Communicate.
Form coalitions.
Negotiate.
Sounds simple? It’s not. Right now, none of the 13 existing protocols tick all these boxes. But some are getting closer than others.
MCP Protocol
Let’s cut straight to the chase. MCP (Model Context Protocol) is the rising star here. Introduced by Anthropic in November 2024, it’s an open standard and open-source framework designed to make large language models (LLMs) and similar systems speak the same language with external tools, systems, and data sources.
Its adoption rate? The biggest so far, which says something about its real-world utility.
Two Main Camps of Protocols
When we talk “AI communication,” protocols tend to fall into two buckets:
Context-oriented
Inter-agent
Context-Oriented Protocols
These are about giving the agent the context it needs to achieve its goal. Think of it as a request-and-response setup. Calling an agent in this scenario is like ordering a coffee: you tell it exactly what you want, it hands it back. No small talk.
Inter-Agent Protocols
Here, agents talk to each other but it’s more than just question and answer. They invoke each other, working on shared or connected goals.
In other words:
Calling is context-oriented.
Invoking is inter-agent.
In context-oriented setups, interactive tools are low-autonomy agents. In agent-to-agent interactions, agents can also be seen as high-autonomy tools, built to accomplish more complex tasks.
***
Meet Some of the Players
Agent.json - Think of this as the agent’s “about me” page, sitting on your domain. It describes the API, explains what it does, and how to use it. Even better — it shows you how to chain calls to get something meaningful done.
A2A Protocol - Agent-to-agent. Calls are expected to take time, so they’re asynchronous.
ANP - Agent Network Protocol. Same idea as A2A — but on the blockchain.
ACP (Agent Connect Protocol) - From Cisco. MCP is running under the hood here. It defines how to find, describe, run, and retrieve an agent — and even download it to host and launch directly.
ACP (Agent Communication Protocol) - From IBM. Similar to A2A. Slightly confusing naming, yes.
Agora - LLM-first protocol. Designed to update itself to achieve its goal. Still in the “idea” stage.
Agent Protocol - Exists… and that’s pretty much all we know.
So, Which One Wins?
Honestly, none. At least not yet. Each protocol covers part of the vision, but the full dream (agents discovering, authenticating, communicating, forming coalitions, and negotiating) is still out of reach.
Still, MCP seems to have the momentum, and if you squint a little, you can see a future where the pieces finally fit.
Disclaimer
I’m not an expert on AI protocols. I’m learning as I go, trying to read details, listening to SME speeches/podcasts, watching videos from conferences, and writing about what I find. If you spot something off or have extra insights, I’m all ears. The text is AI rewritten from personal notes.
