What is A2A? A2A, the Agent2Agent protocol, is an open standard that lets independent AI agents find each other, hand off tasks, and share results, the way MCP lets a single agent connect to tools and data. Google created it in April 2025, gave it to the Linux Foundation that June, and on Monday it joined the Agentic AI Foundation, the same body that has housed MCP since December.
That last move is the news. It sounds like bureaucracy, a protocol changing which committee it reports to. What it actually means is that the two standards most likely to define how AI systems work together are now governed in the same place by the same members, and the list of members is everyone: Amazon, Anthropic, Google, Microsoft, OpenAI, Cloudflare, Block, Bloomberg, and some 250 others. The plumbing of the agent era just got a building code.
What does MCP do, and what does it not do?
MCP connects one agent to the tools and data it needs. That is the whole job. We covered it in What Is MCP, and How Is It Different from an API?: an MCP server wraps something, a calendar, a database, a CRM, and describes it in a standard shape so that any AI that speaks MCP can use it without custom integration code. People call it the USB-C of AI, and the analogy holds. One plug, many devices.
What MCP does not do is let two agents talk to each other as agents. It is built for a single brain reaching out to dumb tools. The tool does not have opinions. It does not plan. It does what it is told and returns a result. That is exactly right for a calendar, and exactly wrong for a second AI system that has its own instructions, its own memory, and its own idea of how to do the job.
MCP is also enormous now. Its two main software kits pass half a billion downloads a month, the specification got a major revision on July 28 that made it simpler to run at scale, and the foundation counted more than ten thousand published MCP servers back when it formed. If you have used an AI assistant that could suddenly read your email this year, MCP is almost certainly the reason.
What does A2A do that MCP cannot?
A2A lets agents treat each other as colleagues rather than as tools. The official description is that it lets independent agents, including ones that use MCP internally, discover each other, delegate tasks, and share results.
The practical difference is in what gets passed back and forth. With MCP, an agent sends a precise instruction and gets a precise answer. With A2A, an agent sends a task, something closer to how you would brief a person, and the other agent works on it for however long it takes, reports progress, asks for clarification if it needs to, and hands back a result. There is a way for an agent to publish what it is good at, so others can find it. There is a way to run a task that takes hours rather than seconds. There is a way to say, partway through, that something has changed.
If MCP is a hand reaching for a tool, A2A is a conversation between two people who each have their own hands.
| MCP | A2A | |
|---|---|---|
| Connects | An agent to tools and data | Agents to other agents |
| Created by | Anthropic, November 2024 | Google, April 2025 |
| Governed by | Agentic AI Foundation (since December 2025) | Agentic AI Foundation (since August 17, 2026) |
| Unit of work | A tool call that returns a result | A task that can run for hours and report progress |
| Typical example | Reading your calendar, querying a database | One vendor’s agent handing a finished job to another’s |
| Current spec | 2026-07-28 revision | Version 1.0, March 2026 |
Why do you need both?
Because the two protocols sit at different layers, and a real agent system needs both layers. The foundation itself described it that way this week: MCP for connecting an agent to tools, A2A for connecting agents to agents, and neither replacing the other.
Picture a small accounting firm a year from now. It runs a bookkeeping agent that reconciles transactions, and a client-communication agent that drafts the monthly summary email. The bookkeeping agent uses MCP to reach QuickBooks and the bank feed. The communication agent uses MCP to reach the email system and the client list. When the books close, the bookkeeping agent uses A2A to hand the communication agent the numbers and say, in effect, write the summary. The second agent may come from a different vendor entirely. It does not matter, as long as both speak the standard.
Without A2A, the firm would have to glue those two agents together with custom code, or buy both from the same vendor and hope. Without MCP, each agent would need custom code for every tool it touches. The two standards together mean you can assemble a system out of parts, the way you assemble an office out of a phone system, a printer, and a laptop from three different companies.
What does this mean if you are buying an agent?
Ask two questions of any agent vendor: does it speak MCP, and does it speak A2A. The first is table stakes in 2026. The second is early, but it is where the market is going, and a vendor who has not heard of it in August 2026 is a vendor who is not paying attention.
The reason to care is lock-in. An agent that only talks to tools through its own proprietary connectors, and only talks to other agents from the same vendor, is a walled garden. The whole point of this week’s news is that the walls are coming down, and the biggest companies in the industry agreed to take them down together. A business that buys into the standards can swap a vendor without rebuilding the office. A business that buys into a garden cannot.
We have written about this pattern before, in The Model Is Not the Product: the intelligence is rented and interchangeable, and the durable value lives in the connections around it. MCP and A2A are what make the connections interchangeable too.
What still is not settled?
Two things, and both are about money and trust rather than plumbing.
The first is payments. Google’s Agent Payments Protocol, the one for letting an agent buy something on your behalf, did not come to the Agentic AI Foundation. Google handed it to the FIDO Alliance, the group behind passkeys, in April. So the standard for how agents talk is in one place and the standard for how they pay is in another. That is probably fine, since paying and talking are different problems, but it means the building code is not one document yet. We looked at the wallet side of this in Agents With Wallets.
The second is identity. When an agent from your vendor hands a task to an agent from another vendor, how does the second one know the first is allowed to ask? A2A has the beginnings of an answer, but the identity companies and the standards bodies are still working out who is on the hook when an agent that should not have been trusted was. That is the next building code, and it is being written right now.
For a business owner, the takeaway is simpler than any of that. Two years ago, connecting AI to your tools meant custom code. One year ago, MCP made it a plug. This week, connecting AI systems to each other became a plug too. The parts are becoming standard. Buy parts, not gardens.
Frequently Asked Questions
What is the Agent2Agent (A2A) protocol?
A2A is an open standard for communication between independent AI agents. It lets agents discover each other’s capabilities, delegate tasks, report progress on long-running work, and share results, regardless of which vendor built each agent. Google created it in April 2025, moved it to the Linux Foundation in June 2025, and it joined the Agentic AI Foundation on August 17, 2026.
What is the difference between A2A and MCP?
MCP connects a single AI agent to tools and data such as a calendar, a database, or a CRM. A2A connects AI agents to each other so they can hand off tasks and collaborate. MCP is agent-to-tool; A2A is agent-to-agent. A complete agent system typically uses both.
What is the Agentic AI Foundation?
The Agentic AI Foundation is a Linux Foundation body formed in December 2025 to govern open standards for AI agents. Its founding projects were MCP, Block’s goose, and OpenAI’s AGENTS.md, and A2A joined in August 2026. Founding platinum members include AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI, and membership has grown past 250 organizations.
Does my business need to care about A2A?
Not for building anything yourself, but yes when choosing vendors. An agent product that supports MCP and A2A can be combined with tools and agents from other companies, which limits vendor lock-in. Asking a vendor whether their agent speaks both standards is a quick test of whether the product is built for the open ecosystem or a walled garden.
Who governs agent payments?
Google’s Agent Payments Protocol was donated to the FIDO Alliance in April 2026, not to the Agentic AI Foundation. Agent communication standards (MCP and A2A) and agent payment standards are governed by different bodies as of August 2026.