The future of AI might be decided not by who builds the smartest agent, but by who gets them talking. Recently, Google quietly kicked off what many are already calling a “protocol war” by launching A2A (Agent-to-Agent), its open standard for AI agent communication. The move comes just days after OpenAI publicly adopted Anthropic’s MCP (Model Context Protocol), which has rapidly become the go-to for tool integration with AI models.

So now, the internet has two new AI protocols, two massive companies backing them, and one giant question: Do we really need both?

What is Google’s A2A?

Google says A2A is about creating a “multi-agent ecosystem across siloed data systems,” and that agents must be able to discover each other and coordinate actions in real time. A2A defines how they do that—no matter who built them.

Announced on April 9, Google’s A2A is designed to help AI agents communicate and collaborate with one another across vendors and platforms. Think of it as WhatsApp for bots. Agents using A2A share a “public card” with their capabilities, host info, and version. They can then interact using methods like server-sent events (SSE), push notifications, or classic request-response.

What about Anthropic’s MCP?

MCP focuses on a different layer of AI interaction. It standardizes how AI models (especially LLMs like Claude or ChatGPT) connect with tools, data sources, and environments. If A2A is agents talking to each other, MCP is agents talking to everything else.

MCP uses a client-server model. Tools and databases expose themselves through MCP servers, while AI models access them via hosts or clients. It’s already becoming a standard in AI integrations, with backing from OpenAI and other LLM players.

Clash or complement?

Still, not everyone is convinced. Solomon Hykes, CEO of Dagger and ex-Docker, tweeted, “In theory they can coexist. In practice, I foresee a tug of war.” He’s not alone. The line between tools and agents is already blurry. Tools are getting smarter, agents are acting more like tools, and developers only have so much energy to build for multiple ecosystems.

Interestingly, Google says A2A “complements” MCP rather than competes with it. In the docs, Google gives the example of a car repair shop: MCP would help an AI agent control the wrench and lift, while A2A would let agents talk to each other or to humans about what’s wrong with the car.

ANP: The third protocol?

While A2A and MCP battle for dominance, there’s also ANP (Agent Networking Protocol), a more community-driven effort that leans heavily on semantic web tech and decentralized identity (DID). It hasn’t gotten the corporate muscle of Google or Anthropic, but it has passionate developers and a flexible, open architecture.

One interesting difference: A2A and MCP mostly use JSON, while ANP uses JSON-LD and schema.org to make agent descriptions easier for other systems to understand.

So who wins?

But with OpenAI already onboard the MCP train, and Google quietly hedging bets by supporting both, this might not be a war with one winner. Instead, we could end up with a layered AI internet where agents talk through A2A and think through MCP.

That depends. Simplicity and adoption usually win in tech. JSON beat XML. HTTP beat SOAP. If A2A ends up easier to use than MCP, developers might flock to it—regardless of technical merits.

Read More:

OpenAI wants to take on X and Instagram with a social app