A2A (Agent-to-Agent) Protocol in Google ADK

Think of it like this: Just as HTTP allows websites to talk to each other, A2A allows AI agents to talk to each other.

Key Concepts

Concept	Description
A2A Server	Exposes your agent so others can call it over the network
RemoteA2aAgent	Client proxy that lets your agent call remote agents
Agent Card	JSON file describing an agent's capabilities (like a business card)
Task	A unit of work exchanged between agents

When to Use A2A vs Local Sub-Agents

Use A2A When...	Use Local Sub-Agents When...
Agents are separate services	Agents share the same process
Different teams maintain agents	Internal code organization
Cross-language communication needed	Performance-critical operations
Formal API contracts required	Simple helper functions

How A2A Works - Architecture Diagram

A2A Communication Flow

Simple Example

Step 1: Create the Remote Agent (to be exposed)

# remote_agent.py
from google.adk.agents import Agent
from google.adk.a2a.utils.agent_to_a2a import to_a2a

# Define a simple agent with a tool
def get_weather(city: str) -> str:
    """Get weather for a city."""
    return f"Weather in {city}: Sunny, 25°C"

remote_agent = Agent(
    model='gemini-2.0-flash',
    name='weather_agent',
    description='An agent that provides weather information',
    tools=[get_weather]
)

# Convert to A2A server (auto-generates agent card)
a2a_app = to_a2a(remote_agent, port=8001)

Run it:

uvicorn remote_agent:a2a_app --host localhost --port 8001

Step 2: Create the Consumer Agent

# main_agent.py
from google.adk.agents import Agent
from google.adk.agents.remote_a2a_agent import RemoteA2aAgent

# Create a proxy to the remote weather agent
weather_agent = RemoteA2aAgent(
    name='weather_service',
    agent_card_url='http://localhost:8001/.well-known/agent.json'
)

# Your main agent that uses the remote agent
root_agent = Agent(
    model='gemini-2.0-flash',
    name='assistant',
    description='A helpful assistant',
    sub_agents=[weather_agent]  # Add remote agent as sub-agent
)

Run it:

adk web .

Step 3: Interact

User: What's the weather in Tokyo?

Bot: Let me check the weather for you.
     Weather in Tokyo: Sunny, 25°C

Agent Card Structure

The Agent Card is auto-generated or manually created. It lives at /.well-known/agent.json:

{
  "name": "weather_agent",
  "description": "An agent that provides weather information",
  "url": "http://localhost:8001",
  "version": "1.0.0",
  "protocolVersion": "0.2.6",
  "skills": [
    {
      "id": "get_weather",
      "name": "get_weather",
      "description": "Get weather for a city"
    }
  ],
  "defaultInputModes": ["text/plain"],
  "defaultOutputModes": ["text/plain"]
}

Multi-Agent System Example

Quick Reference

Action	Code
Expose agent	`a2a_app = to_a2a(agent, port=8001)`
Run A2A server	`uvicorn module:a2a_app --port 8001`
Connect to remote	`RemoteA2aAgent(agent_card_url='...')`
Check agent card	Visit `http://host:port/.well-known/agent.json`

Summary

A2A = Protocol for agents to communicate over network
Expose your agent using to_a2a() → creates A2A Server
Consume remote agents using RemoteA2aAgent
Agent Card = JSON describing capabilities (auto-generated or manual)
Works across different frameworks, languages, and machines