What Google's A2A Joining the Agentic AI Foundation Means for Enterprise Agent Architecture

(© Turgay Koca - Canva) In The Hitchhiker's Guide to the Galaxy, a race of hyper-intelligent pan-dimensional beings build a supercomputer to work out the answer to Life, the Universe, and Everything, wait seven and a half million years, and get 42 - at which point they uncomfortably realize that they never quite pinned down the Question. Enterprise architecture has been running an eerily similar experiment on itself during 2026. The computers may be faster and the wait is shorter, but it still involves a great deal of expensive machinery, an enormous quantity of tokens, and a question that somehow is always slightly under-specified. It usually gets phrased as "what's the ROI on our agentic AI?" - which, as any architect will tell you after their second coffee, is really several questions in a trenchcoat. Mazin Gilbert, Executive Director of the Agentic AI Foundation (AAIF), has a coherent answer to a well-specified version of that question. When we spoke shortly after Google's Agent2Agent Protocol (A2A) joined the AAIF as its fifth hosted project, alongside Model Context Protocol (MCP), goose, Agents.md and agentgateway, he highlighted a piece of open infrastructure that hasn't received as much airtime as it probably should. For some additional context, the AAIF is the Linux Foundation body that houses the open protocols and reference implementations underneath enterprise agent systems. Under the AAIF's model, a hosted project is where the foundation stewards the specification and provides neutral governance, so the project keeps its own maintainers and technical direction, but the specification evolves under the AAIF's umbrella. Google's contribution adds the layer that lets independent agents discover each other and hand off work. Even though Google donated the specification to the Linux Foundation in 2025, the coordination within the same governance family is an aspect I was interested in getting into. What A2A does, and why it matters MCP and A2A solve different problems. MCP is how an agent connects to the outside tools, APIs, databases or files it needs to do its job. Without it, Mazin notes, an agent is "dumb". It has been in production for long enough to be a familiar part of enterprise agent systems - Anthropic's adoption figures show around half a billion downloads a month. A2A lets an agent talk to another agent in an autonomous system that has its own agenda, security set-up and maintainer. Mazin explained: > Imagine that my agents need to make a transaction and need to speak to a merchant's agents - Safeway, Marks and Spencer, whatever the case may be. They are not both my agents, so I don't have the complete control. My agents are talking to external agents. That's A2A. This enables two agents to discover themselves. First you have to know whether they exist, and then they have to transact with themselves in a very secure way. The AAIF's own materials describe it as agent-to-agent interoperability across framework and vendor boundaries. Mazin describes the bigger picture as: > You need to create this mesh network of protocols, the foundation, the plumbing that's interoperable, and that's true for the Internet of Information today, of humans today. You have to do the same thing here. And it's not one protocol. it's not five protocols. It's a number of them that have to come together to complete this puzzle. The pattern APAC has already been building One aspect of our discussion covered where A2A demand was coming from. Mazin said that in the US and Europe, the enterprise agents story has been shaped by data center economics, with MCP as the connective tissue between agents and the tools, APIs and data they need to reach. Western enterprise readers have been following the topic for some time, but Mazin explained that a different pattern has been forming: > When you get to APAC, the story is really the consumer story. They start from the consumer, they start from the end device. So imagine your assistants on your phone, on the appliances in your home, your cars and connected cars - and every single one has their own agents. My TV cannot talk to my car or somebody else's, you can't do that without a secure, trustable protocol between the two. And this is A2A. AAIF's materials show how far along this has already developed. Huawei has standardized A2A as the protocol between Celia, its OS-level AI assistant, and in-app agents across the Harmony OS developer platform - covering scenarios like Celia handing off long-running tasks to an app agent, or requesting contextual recommendations from an application agent. Tencent's WeChat is among the first major apps integrating with Huawei and other Android OEM assistants over A2A, enabling messages, voice calls and video calls initiated through an AI assistant, running with dual authorization through the protocol. Requests for A2A support, Mazin notes, have come predominantly from that region. For Western enterprise readers, the pattern is coming back - on August 13th, four days before the A2A move into the AAIF was announced, the foundation added 57 new members including Visa, Wells Fargo and Alibaba. Mazin elaborated: > You're going to see this beginning to change even in the US and Europe. B2B2C is emerging, and financial companies are starting to play a big role. Within three years from now, it's all going to become mainstream, enterprise, consumer, everything. Next on the agenda - trust and scale Over the next six to twelve months, Mazin says there are two key areas of focus for AAIF work. One is scale - the operational reality of moving agent systems from pilot into production with reliability, day-two operations and load-bearing engineering that keeps things running. The other is trust. He observes: > Agents are going to make mistakes. Agents hallucinate. Who is accountable? > > > You. The consumer, the enterprise. It cannot be the agent. Trust encompasses guardrails, auditability, traceability and accountability so that an enterprise can stand behind an agent's actions. The AAIF already has named working groups, and eight have been formally constituted: Accuracy & Reliability, Agentic Commerce, Governance/Risk/Regulatory Alignment, Identity & Trust, Observability & Traceability, Security & Privacy, Workflows & Process Integration, and Taxonomy & Landscape. Between them, they cover active streams each with their own charters, member organizations and deliverables. As an example, the Agentic Commerce group ties directly back to the AP2 and A2A work - transactions between agents need trust and identify primitives that don't yet have industry-wide answers, and it's going to take some work. On July 28th, MCP went through what the foundation described as its largest revision since launch with a move to a fully stateless architecture that lets MCP servers run behind conventional load balancers and Kubernetes clusters, an OAuth hardening pass that closed a class of authorization vulnerabilities, and a formal 12-month deprecation guarantee. That working group cycle is an example of what A2A - currently backed by around 150 organizations - needs to grow into. Admittedly, A2A is early. But there is a governance shape which already has trust-and-scale work under construction. My take So what happens next? AGNTCon + MCPCon Europe is on 17-18 September in Amsterdam, which is where I'll be attending to find out more.