Agent-to-Agent Protocol (A2A)

Overview

The Agent-to-Agent (A2A) Protocol enables AI agents to discover each other, negotiate capabilities, and delegate tasks—regardless of their underlying framework or vendor.

If MCP is how agents talk to tools, A2A is how agents talk to each other.

Core Concepts

Agent Cards

Agent Cards are JSON documents that describe an agent's capabilities:

{
  "name": "Research Agent",
  "description": "Searches and synthesizes information",
  "skills": ["web-search", "summarization"],
  "endpoint": "https://agent.example.com/a2a",
  "authentication": { "type": "oauth2" }
}

Agent Cards are discoverable at /.well-known/agent-card.json.

Task Lifecycle

1. Discovery: Find agents via Agent Cards
2. Negotiation: Agree on task parameters
3. Execution: Remote agent processes task
4. Streaming: Real-time progress updates
5. Completion: Results with artifacts

Multimodal Support

A2A handles rich content: text, images, video, interactive forms, and custom UI widgets.

Technical Details

• Transport: HTTP/HTTPS, gRPC (v0.3+), SSE, WebSocket
• Format: JSON-RPC 2.0
• Auth: OAuth 2.0, API Keys, Signed Agent Cards, mTLS

IBM ACP Merger

In August 2025, IBM's Agent Communication Protocol (ACP) merged with A2A under the Linux Foundation. This unified the two leading standards, with IBM joining the A2A Technical Steering Committee.

IETF Standardization

Several IETF drafts are building on A2A:

• ANS (Agent Name Service): DNS-based agent discovery with PKI
• DN-ANR: DNS SVCB records for agent protocols
• BANDAID: Structured DNS namespace for AI discovery

Enterprise Adoption

150+ organizations support A2A including:

• Microsoft (Azure AI Foundry, Copilot Studio)
• SAP (Joule integration)
• Salesforce, ServiceNow, Workday

Real deployments: Tyson Foods, Gordon Food Service (supply chain agents).

Relationship with MCP

Aspect	MCP	A2A
Primary Use	Agent → Tool	Agent → Agent
Architecture	Client-Server	Peer-to-Peer
Focus	Tool invocation	Task delegation

Most production systems use both.

When to Use A2A

Use A2A for:

• Multi-agent orchestration
• Cross-vendor collaboration
• Enterprise workflows
• Agent marketplaces

Don't use A2A for:

• Simple tool calling (use MCP)
• Decentralized discovery (use ANP)

# Python A2A Client Example
from a2a import A2AClient, AgentCard

# Discover remote agent via well-known path
card = AgentCard.fetch(
    "https://agent.example.com/.well-known/agent-card.json"
)

# Create client and send task
client = A2AClient(card)
task = await client.create_task(
    skill="summarize",
    input={"document": "...long text..."},
    stream=True
)

# Stream results
async for event in task.stream():
    if event.type == "progress":
        print(f"Progress: {event.progress}%")
    elif event.type == "complete":
        print(f"Result: {event.result}")