Table of Contents
Now that AI agents are commonplace, the next challenge has shifted to "how to make agents collaborate with each other." If MCP is the standard that connects an agent to its "tools," then A2A (Agent2Agent) is the standard that connects an agent to "another agent." It lets AIs built on different vendors and frameworks talk and cooperate through a common convention.
This article lays out, for beginners, what A2A is, why it's needed, how it differs from MCP, how Agent Cards and Tasks work, and where it stands plus implementation.
MCP is vertical (tools), A2A is horizontal (peers)
— AIs talk and cooperate through a shared convention
MCP = vertical connection
Connect one agent to tools and data.
A2A = horizontal coordination
Connect agents to each other and delegate work.
1. What Is A2A?
A2A (Agent2Agent) is an open standard (protocol) for AI agents to discover, communicate with, and cooperate with each other across the frameworks they were built on. Google released it in April 2025, donated it to the Linux Foundation that June, and it reached v1.0 in 2026. It's being shaped into a "common language" not tied to any single company.
Picture "the etiquette of a business partnership between companies." If MCP is your own staff using tools, A2A is the shared rulebook for asking another company (another AI) to do a job and receiving the result. Whatever the other side is built with, you can exchange "please do this" and "done" through a fixed procedure.
💡 In one line: A2A = "a common conversation protocol between AI agents." Where MCP handles "connecting to tools," A2A handles "coordinating with peers."
2. Why It's Needed: The Era of Agent Collaboration
In 2026, AI is heading toward a model where, instead of one agent doing everything, multiple agents with different specialties divide the work and cooperate. For example, a request like "arrange a trip" gets handled as a relay: a planning agent → a hotel-booking agent → a payment agent.
But when each agent is built with a different vendor and technology, connecting them takes custom work every time. That's where a common protocol — A2A pays off. As long as everyone follows the shared convention, you can turn multi-agent systems into "parts you can combine." It's the same shift that made MCP the common standard for tool connections.
3. How It Differs from MCP (Vertical vs. Horizontal)
A2A and MCP aren't competitors — they have different roles, often explained as "vertical vs. horizontal." Stacking both is the 2026 standard setup.
Agent ↔ tools
Connect one agent to tools and data — a DB, an API, files. A connection that "adds capability."
Agent ↔ agent
Connect agents to each other and delegate work back and forth. Coordination that "cooperates with peers."
The mnemonic is simple: "connect to tools = MCP; connect to peers = A2A." In a real system, each agent holds its own tools via MCP while coordinating with other agents via A2A — this vertical-MCP-plus-horizontal-A2A two-layer setup is becoming the default route for enterprise agent operations.
4. How It Works: Agent Cards and Tasks
The heart of A2A is the "Agent Card." It's a JSON like a "business card" that each agent publishes, stating "what I can do," "where to talk to me," and "how authentication works." Its location is even fixed at /.well-known/agent-card.json.
Discover (Agent Card)
Read the other agent's card and grasp "what it can do."
Request (Task)
Send a "Task." It carries states — working, input-required, completed.
Result (Artifact)
Receive the output. Long-running tasks can stream progress too.
Communication is built on widely used technology — HTTP, Server-Sent Events (SSE), and JSON-RPC 2.0. The key point is that agents don't expose their internals to each other. Each one keeps its contents (the tools it uses, its reasoning) hidden and exchanges only tasks and results. That's why they can collaborate safely even across companies.
⚠️ Security is still a must: agent collaboration is convenient, but you need a design that doesn't over-trust external agents. Combine it with guardrails and permission management.
5. Where It Stands, and Implementation
A2A is spreading fast. Per the Linux Foundation's announcement (as of April 2026), more than 150 organizations use it in production, it has 22,000+ GitHub stars, and SDKs ship in five languages (Python, JavaScript, Java, Go, and .NET). Majors like Microsoft, Salesforce, SAP, and ServiceNow are involved, cementing its place as an industry standard.
On implementation, the per-language SDKs make it fairly easy to build both "the side that publishes an Agent Card (the server)" and "the side that delegates to other agents (the client)." The recommended order is to first build a single agent, then, once comfortable, have it coordinate with others via A2A. Like Skills and MCP, the foundation is the idea of "componentizing through a common format."
※ Figures and specifications are cited from various disclosures (as of June 2026). The protocol is evolving; check the Linux Foundation / official spec for the latest.
Summary
Three takeaways on A2A.
- What it is: an open standard for AI agents to discover and collaborate across frameworks. Born at Google, governed by the Linux Foundation.
- The difference: MCP is vertical (agent ↔ tools); A2A is horizontal (agent ↔ agent). Stacking both is the standard.
- How it works: discover via an Agent Card (business card) → Task (request) → Artifact (result). Cooperate safely while keeping internals hidden.
From one agent to a "team" that cooperates — A2A is the bridge. Read it alongside MCP and multi-agent systems to grasp the full picture of agent collaboration in 2026.
FAQ
Q. Should I use A2A or MCP?
A. It's not either/or. To connect an agent to tools and data, use MCP; to make agents coordinate, use A2A. In a real system you stack both: "each agent holds its tools via MCP and cooperates via A2A."
Q. What's an Agent Card?
A. An agent's "business card." It's a JSON stating "what it can do, where to talk to it, and how to authenticate," placed at /.well-known/agent-card.json. The other side reads it to decide whether it can delegate.
Q. Can it coordinate with another company's AI?
A. That's the whole aim of A2A. Because it's a common protocol, agents from different vendors and frameworks can coordinate. Each keeps its internals hidden and exchanges only tasks and results.
Q. Do solo developers need A2A?
A. If you have only one agent, no. A2A pays off when you want several independent agents to coordinate. It's enough to build a single agent first and adopt A2A once coordination becomes necessary.
