Google A2A — Agent-to-Agent Communication Protocol

How A2A Works

The Problem

MCP connects agents to tools. But what about agent-to-agent communication?

Without A2A:
  Agent A (LangGraph) -- cannot talk to -- Agent B (CrewAI)

With A2A:
  Agent A (any framework) <-- A2A Protocol --> Agent B (any framework)

Agent Cards (Discovery)

Every A2A agent publishes an Agent Card at /.well-known/agent.json:

{
  "name": "security-auditor",
  "description": "Scans code for security vulnerabilities",
  "capabilities": ["security_audit", "dependency_scan", "owasp_check"],
  "endpoint": "https://security-agent.example.com/a2a",
  "authentication": { "type": "bearer" }
}

Other agents discover capabilities by fetching Agent Cards.

Task Delegation

from a2a import A2AClient

client = A2AClient("https://security-agent.example.com/a2a")

# Send a task to the security agent
result = await client.send_task({
    "type": "security_audit",
    "input": {"code": "...", "language": "python"}
})
# result: {"vulnerabilities": [...], "risk_score": "low"}

Streaming Responses

async for chunk in client.stream_task({"type": "code_review", "input": {...}}):
    print(chunk)  # Real-time updates from the remote agent

Multi-Agent Architecture

Orchestrator Agent
  |-- A2A --> Code Writer Agent (writes code)
  |-- A2A --> Security Agent (reviews for vulnerabilities)
  |-- A2A --> Test Agent (generates and runs tests)
  |-- A2A --> Deploy Agent (deploys to production)

A2A vs MCP

They are complementary:

• MCP: Agent uses a GitHub tool to create a PR
• A2A: Agent asks another agent to review the PR

Key Stats

• 10,000+ GitHub stars
• By Google DeepMind
• Agent discovery via Agent Cards
• Task delegation with streaming
• Framework-agnostic

FAQ

Q: What is A2A? A: An open protocol by Google for AI agents to discover and communicate with each other, enabling multi-agent systems across different frameworks.

Q: Is A2A free? A: Yes, open protocol and SDK under Apache 2.0.

Q: A2A vs MCP? A: MCP connects agents to tools. A2A connects agents to other agents. Use both together for full agent ecosystems.