Here is a quick overview of the various protocols we hear and read so much recently in the AI space. I don’t intend to go into the details, just a very brief and human readable, objective view.
MCP
Let’s start with MCP (Model Context Protocol), which I covered in one of my recent blog posts. Agents need tools to interact with their environment and MCP aims there: provide a framework for writing and integrating tools into our agents.
- provide reusable tools
- protocol to integrate tools into compatible framework
A2A
A2A (Agent2Agent) aims at agent to agent communication. Therefore, the focus is on how one agent interacts with another agent to enable complex multi agent communication. Through agent cards, the agents capabilities are discoverable, think of the agent card as the OpenAPI specification of your agent.
- agent to agent communication protocol
- sits on top of your agent framework, is framework independent
- agent cards provide a means of describing the agent’s capabilities
ACP
ACP (Agent Communication Protocol) is also aiming at standardizing agent to agent communication. It prioritizes asynchronous communication to support long running tasks and uses simple REST APIs
- agent to agent communication protocol
- sits on top of your agent framework, is framework independent
- A2A and ACP have a very similar programming model
And now?
From my perspective, it is a positive development that the industry is trying to standardize the communication protocols for tools and agent to agent interactions. In practice, agent frameworks will need to support the protocols to make them useful. See for example how Langgraph supports MCP tools.
There is a bit of a hype around MCP, A2A, ACP and most likely, whatever comes next. If we look at it from a more technical standpoint, they are enabling technologies and relatively simple in nature. I really recommend trying them out to gain some hands on experience.
Edifying