Agent2Agent Protocol (A2A) - Google's AI Agent Protocol

Published on

This week, Google made another important move on the AI AGENTS chessboard: the Agent2Agent Protocol (A2A).

When I first read about this launch, right after the MCP “boom” from Anthropic in the last few weeks, I confess I had that feeling from Marcus Mendes’s joke about plug standards - “We have 14 competing standards!” -> “That’s ridiculous, we need to create ONE universal standard!” -> “Situation: now we have 15 competing standards.” 😂

But after looking at it more carefully, even though it can look like a direct competitor, I saw that it can, in fact, work as a complement to MCP (at least for now). It’s also worth highlighting that both Demis Hassabis and Sundar Pichai publicly stated they’re going to integrate MCP into Google’s tools, showing that the two protocols can coexist perfectly.

So, what’s the fundamental difference today?

📑 MCP (Model Context Protocol): Connects an Agent to tools/resources (web search, maps, etc.) in a standardized way.
📑 A2A (Agent2Agent Protocol): Lets different Agents talk and collaborate with each other, even if they come from different vendors or frameworks, without necessarily sharing memory or tools directly.

It’s as if MCP were responsible for giving tools to a worker, and A2A let workers specialized in different areas collaborate.

Imagine you have an Agent A that needs a tool it doesn’t have, but knows that another agent (Agent B) is a pro at it. Through A2A, Agent A can “ask Agent B for help”, and then both can individually use MCP to access the tools they need.

In practice, with A2A, each agent describes itself using an “Agent Card” (a JSON that says who it is, what it does, etc.), allowing other agents to discover how to interact with it.

And here comes an interesting possibility that I discussed in the last episode of the “IA Sob Controle” podcast: in the future, maybe A2A will become more central than MCP in many cases.

Think about it: today you can connect your agent to the GitHub “tool” (for example) via MCP. But what if GitHub launches its own agent, super optimized for code/repository tasks? In that scenario, it might make more sense for your agent to talk directly to GitHub’s agent via A2A, which is more specialized, reducing the need for the MCP connection to GitHub’s (more “general”) tool in that specific context. It’s a possibility worth watching!

A2A is an open-source project and Google wants the community to contribute. It’s still VERY early, but it’s something to keep an eye on, since it has the potential to become an important standard in building multi-agent systems.

If you work with AI development, it’s well worth taking a look at the A2A documentation. The code is already available on GitHub, and I’m leaving the links below! 😉