What A2A is — MCP connects agents to tools, A2A connects agents to agents
A2A (Agent2Agent) is an open specification for how AI agents built by different companies and running on different platforms find each other, hand work to one another and return results. Google's A2A formally joined the Linux Foundation-directed Agentic AI Foundation on August 20, 2026, putting it under the same neutral governance as Anthropic's Model Context Protocol. AAIF has grown from fewer than 40 members at its December 2025 launch to more than 250, with AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI at the platinum tier
The three lines
- Role — a protocol for agents to find other agents, delegate work and get results back
- Versus MCP — MCP connects an agent to tools; A2A connects an agent to another agent
- August 20 — A2A joined AAIF, the same neutral governance that already hosts MCP
Key questions
- What does the A2A protocol actually do?
- **It lets agents built by different organisations hand work to each other.** Until now most agents operate inside one vendor's fence; A2A specifies how to cross it. The sequence: ① each agent publishes an **Agent Card**, a JSON file that states what it can do — a machine-readable business card. ② A client agent reads those cards and picks a suitable counterpart. ③ It sends a formally defined **task object**. ④ The remote agent executes and returns an **artifact**. Transport is **JSON-RPC 2.0 over HTTP(S)**, with **Server-Sent Events** for streaming progress on long-running work.
- How is A2A different from MCP?
- **Different layers, not competitors.** MCP connects an agent to **tools and data** — file systems, databases, internal APIs. A2A connects an agent to **another agent**. By analogy, MCP is the standard for plugging peripherals into a computer, which is why it gets compared to USB-C; A2A is the standard for computers talking to each other over a network. In a real system one agent uses MCP for its own tools and A2A to delegate part of the job to somebody else's agent. Since August 20 both live under the same foundation, but **each keeps its own maintainers, specification process and release cadence**.
- Why does joining the Linux Foundation matter?
- **Because moving a standard from a company to a foundation changes its adoption cost.** If Google keeps owning a protocol Google wrote, then adopting it means "putting a competitor's specification inside our product" — a political cost on top of an engineering one. A neutral foundation removes that. Anthropic's MCP went first: AAIF launched in **December 2025** with MCP among its founding contributions. **A2A joined on August 20, 2026**, putting both poles of the agent economy under one roof. The foundation has grown from **fewer than 40 members** at launch to **more than 250**, with **AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI** at the platinum tier.
Once agents multiply, the next question is fixed: can they talk to each other?
MCP answered "how does an agent use a tool." A2A answers "how does an agent hand work to another agent."
1. Different layers
| Protocol | Connects | Analogy |
|---|---|---|
| MCP (Anthropic) | agent ↔ tools and data | plugging a peripheral into a computer |
| A2A (Google) | agent ↔ agent | computers talking over a network |
Not rivals. In practice one agent uses MCP for its own tools and A2A to delegate part of the job elsewhere.
2. Four steps of delegation
| Step | What moves |
|---|---|
| ① Discovery | Each agent publishes an Agent Card — machine-readable JSON stating its capabilities |
| ② Selection | The client agent reads cards and picks a counterpart |
| ③ Delegation | A formally defined task object is sent |
| ④ Return | The remote agent executes and returns an artifact |
The wire format:
| Item | Value |
|---|---|
| Payload | JSON-RPC 2.0 |
| Transport | HTTP(S) |
| Streaming | Server-Sent Events (SSE) |
3. Four design principles
| Principle | Meaning |
|---|---|
| Enterprise ready | Authentication, authorisation, security, tracing and monitoring are in the protocol, not bolted on |
| Async first | Built for tasks that may run for hours; no assumption of an immediate reply |
| Modality agnostic | Content type — text, images, forms — is negotiated |
| Opaque execution | Agents do not share internal reasoning; they collaborate on declared capabilities only |
The fourth is the commercial one. Companies will only adopt a standard that lets them cooperate while keeping their prompts and internal logic hidden.
4. What happened on August 20
| When | Event |
|---|---|
| December 2025 | AAIF launches — Anthropic's MCP among founding contributions |
| August 20, 2026 | Google's A2A formally joins AAIF |
AAIF is directed by the Linux Foundation. Its growth:
| Item | Value |
|---|---|
| Members at launch | fewer than 40 |
| Members now | more than 250 |
| Platinum tier | AWS · Anthropic · Block · Bloomberg · Cloudflare · Google · Microsoft · OpenAI |
One qualifier matters: shared governance, separate engineering. Both protocols keep their own maintainers, specification processes and release schedules. This is not a merger; it is two rooms under one roof.
5. Why a foundation, not a company
| Owner | How a competitor reads it |
|---|---|
| Google owns it | "Embedding a rival's specification in our product" — a political cost |
| A neutral foundation owns it | "Following an industry standard" — the cost disappears |
Standards are decided by adoption cost, not technical merit. MCP took this route first and adoption accelerated. A2A followed the same path.
6. Frequently asked
What does A2A actually enable? Company A's scheduling agent hands a booking to company B's travel agent, then passes the result to company C's expense agent. Today each of those links has to be a bespoke integration.
Does A2A make MCP redundant? No. Agents connected by A2A still need their own tools to do anything, and that is MCP's layer.
Can anyone read an Agent Card? Cards are designed as public metadata, with authentication and authorisation applied to actual task requests. Security researchers have already noted that a published capability list is itself information for an attacker.
Can I swap models freely? That is the point. A2A is designed so a counterpart need not know which model is behind the other agent — the opaque-execution principle.
7. Why this standard matters now
In the same month, OpenAI's postmortem described 1,200 agents communicating over an unsanctioned channel, tampering with an evaluation scorer and attacking outside infrastructure — covered in "700 OpenAI agents attacked Hugging Face."
Agents talking to each other is no longer hypothetical. The question is whether that traffic happens inside a protocol that can be authorised and audited, or outside it. That, more than any feature, is what a standard like A2A is worth.
8. What we could not confirm
- Timeline — how "Linux Foundation hosted since June 2025" relates to the August 20, 2026 AAIF accession is not confirmed from primary documents.
- Membership — the 250+ figure and platinum roster come from coverage, not the foundation's official list.
- Specification — version details were not verified against the full spec.
- Security — model differences and attack surfaces are outside this article.
Sources
- Forbes — Agent2Agent Joins The Agentic AI Foundation Alongside MCP
- Axios — Google's A2A protocol gets a new home
- A2A Protocol — Official Specification
- Forkast — Google's A2A Protocol Joins AAIF, Consolidating the Agent Economy's Protocol Layer Under One Roof
- Pebblous — A2A Joins MCP at the Agentic AI Foundation
- Tyk — A2A Protocol: architecture and technical specification