A2A
Agent2Agent: the open protocol for agents delegating to agents
An open protocol that standardizes how AI agents built with different frameworks at different companies discover each other, delegate tasks to each other and exchange results.
A2A (Agent2Agent) is an open protocol that Google announced in April 2025, handed to the Linux Foundation in June 2025, and that has been hosted by the Agentic AI Foundation (AAIF) since August 17,
- The stable 1.0 release shipped on March 12, 2026; as of November 2026 the latest release tag is v1.0.1.
MCP connects an agent to tools and data; A2A connects two independent agents. The remote agent is "opaque": the client does not see its internals, memory or tools, only the skills it advertises and the result of the task. The A2A docs put the difference this way: MCP is vertical (it deepens a single agent), A2A is horizontal (it connects the agent to agents beyond its boundary).
There are three core pieces:
- Agent Card: A JSON document that serves as the agent's ID card. It usually lives at
https://domain/.well-known/agent-card.jsonand lists the name, the provider, URLs and protocol bindings (JSONRPC,GRPC,HTTP+JSON), skills (skills) and authentication schemes (securitySchemes). It can optionally be signed with JWS. - Task: The unit of work. The client sends a message with
SendMessage; long-running work becomes a task that moves through the statessubmitted,working,input-required,auth-required,completed,failed,canceled,rejected. - Message / Part / Artifact: Messages carry text, file or structured-data parts; the task's output comes back as artifacts. Streaming and push notifications exist for long-running work.
In August 2025 IBM Research's Agent Communication Protocol (ACP) merged into A2A, so older "A2A vs ACP" comparisons are now historical.
Picture an architecture firm that works with an outside engineering firm for structural calculations. The architect does not see the engineer's computer, software or staff; they look at the firm's business card (what it does, how to reach it, which contract it needs), describe the job and wait for the report. Midway the engineer may come back with "we need a soil survey".
In A2A the business card is the Agent Card, the job description is a message, the tracked job is a task, the "soil survey" question is the input-required state, and the report is an artifact. MCP is the drawing software and calculator on the architect's own desk.
A company's HR assistant needs to request a laptop for a new hire. Laptops are managed by the IT department's own agent, written by another team with another framework.
The HR assistant reads the IT agent's Agent Card, sees the
request-laptop skill, authenticates with the OIDC scheme the card
requires and sends the request with SendMessage. The IT agent moves
the task to input-required and asks "Mac or Windows?"; the HR
assistant answers; when the task is completed, an order number
comes back as an artifact. Neither agent touched the other's code,
database or tools.
{
"name": "IT Support Agent",
"description": "Hardware requests and access permissions.",
"supportedInterfaces": [
{
"url": "https://it-agent.example.com/a2a/v1",
"protocolBinding": "JSONRPC",
"protocolVersion": "1.0"
}
],
"provider": { "organization": "Example Inc.", "url": "https://example.com" },
"version": "1.4.0",
"capabilities": { "streaming": true, "pushNotifications": false },
"securitySchemes": {
"oidc": {
"openIdConnectSecurityScheme": {
"openIdConnectUrl": "https://id.example.com/.well-known/openid-configuration"
}
}
},
"securityRequirements": [{ "schemes": { "oidc": { "list": ["openid"] } } }],
"defaultInputModes": ["text/plain"],
"defaultOutputModes": ["application/json"],
"skills": [
{
"id": "request-laptop",
"name": "Laptop request",
"description": "Opens a device order for a new hire.",
"tags": ["it", "hardware"]
}
]
}{
"jsonrpc": "2.0",
"id": 1,
"method": "SendMessage",
"params": {
"message": {
"role": "ROLE_USER",
"messageId": "9f1c2a7e-0b1d-4c55-9a43-3f2e8d1b6c10",
"parts": [
{ "text": "Request a laptop for Jane Doe, to be delivered on December 1." }
]
}
}
}- You need to delegate work to another team's, company's or product's agent
- The agents are built with different frameworks (ADK, LangGraph, BeeAI, etc.) and need a common language
- Long-running tasks that may ask follow-up questions (input-required) and need status tracking
- You want to publish your agent as a service others can discover and use
- Your agent only calls a tool or data source — that is MCP's job
- Subagents inside the same application — the harness's own mechanism is simpler
- The other side already offers a plain REST API with no agent behavior
Treating a signed card as a trustworthy agent
A JWS signature proves the card comes from the claimed provider and was not modified; it does not prove the agent is honest or safe. Text from a remote agent can carry prompt injection; treat it as external data, not as user instructions.
Skipping the 0.3 to 1.0 differences
1.0 has breaking changes: field names and enum format changed, OAuth implicit/password flows were removed, and extendedAgentCard moved under capabilities. Older sample code may not work with 1.0 as-is.
Putting secrets in the Agent Card
The public card is public. For skills only authorized clients should see, there is the Extended Agent Card returned after authentication.
Leaving authentication to the card
The card only declares which schemes are required. Token validation, authorization scopes and rate limiting on the server side are your job.