Agentic AI Building Blocks: Google’s Protocol Joins Linux Foundation’s AI Group

AI agents built by different companies are beginning to talk to one another, but making that work at scale will require common standards rather than a thicket of custom integrations.

That is the problem Google’s Agent2Agent Protocol, or A2A, is designed to address. Google launched A2A in 2025 as an open protocol allowing independently developed AI agents to discover one another, communicate and delegate tasks across different frameworks and vendors. Google later donated the protocol to the open-source nonprofit Linux Foundation, and A2A is now backed by more than 150 organizations.

A2A recently joined the nonprofit’s Agentic AI Foundation (AAIF), putting it under the same umbrella as Anthropic’s Model Context Protocol, or MCP. The protocols are designed primarily for different jobs. MCP standardizes how agents connect to tools and data, while A2A standardizes how independent agents communicate and collaborate with one another. But companies can use either protocol independently, and their capabilities can overlap in some cases.

For example, A2A is already being used to connect agents across company boundaries. Huawei, for example, has standardized A2A as the protocol between Celia, its HarmonyOS AI assistant, and in-app agents. Tencent’s WeChat is among the first major applications integrating with Huawei and other Android device makers’ assistants through A2A.

The AI Innovator spoke by email with Mazin Gilbert, executive director of the Agentic AI Foundation, about why the industry needs A2A when MCP already exists, whether competing vendors will really allow their agents to work together and what it would take to create what he calls an “internet of agents.”

The AI Innovator: How is A2A different from MCP and why do we need both?

Mazin Gilbert: No single company will build every agent an enterprise needs, nor should it have to. A2A provides a common language for independently developed agents to communicate, even when their internal systems and models are entirely different. MCP often connects an agent to tools and data. A2A enables independent agents to securely collaborate toward a shared outcome without requiring them to expose how they work internally. Both are essential. Not every capability should be modeled as an agent, and not every agent should be modeled only as a capability.

Tech companies all have their own AI platforms. Can their agents truly work together rather than keeping customers inside their own ecosystems?

They can, but interoperability is as much a market commitment as a technical achievement. An enterprise should be able to use a finance agent from one provider to work with a supply chain agent from another while retaining control of its data and workflows.

Neutral governance is critical. A2A evolves around the needs of the ecosystem, not one vendor’s commercial interests. Companies should compete on agent quality and earn customer loyalty through performance, not technical barriers.

You mention the possibility of an “internet of agents.” What has to happen before that becomes real, and how far away are we?

We’re already building the Internet of Agents, with A2A playing a key role in early deployments. Most agent connections today are still pre-planned. The next step is discovery and commerce at scale, where independent agents can discover each other, negotiate and complete a transaction without hand wiring.

I believe adoption will happen quickly. Agent traffic has already exceeded human traffic on the internet. Within the next three to five years, we may have more agents operating in the world than people, which makes clear accountability for their actions increasingly important.