AG2

Overview

AG2 (formerly AutoGen) is an open-source framework for building AI agents. AG2 has built-in support for the A2A protocol through A2aRemoteAgent, allowing your agents to communicate with remote A2A agents like StackOne’s. Use AG2 to build agents that:

Consume StackOne’s A2A agents as remote agents
Orchestrate multi-agent conversations with local and remote agents
Access StackOne platform actions without managing tool definitions

A2aRemoteAgent supports only asynchronous methods - this is a limitation of the A2A client.

Installation

Install AG2 with A2A support using uv. Include your model provider’s extra (OpenAI here), and pin the a2a-sdk 0.3 line:

uv add "ag2[a2a,openai]" "a2a-sdk[http-server]>=0.3,<0.4"

Pin a2a-sdk to the 0.3 line: StackOne’s A2A server implements A2A 0.3, and the current ag2[a2a] otherwise resolves a2a-sdk 1.x (A2A v1.0), which is not compatible yet, so pinning 0.3 selects a compatible AG2 build. Also pass max_turns to a_initiate_chat (below) to bound the conversation, as each turn runs the remote StackOne agent end to end.

Quick Start

This example creates a local agent that delegates HR tasks to a StackOne A2A agent.

import asyncio
import base64
from autogen import ConversableAgent, LLMConfig
from autogen.a2a import A2aRemoteAgent, HttpxClientFactory

# StackOne A2A configuration
STACKONE_API_KEY = "<stackone_api_key>"
STACKONE_ACCOUNT_ID = "<account_id>"
BASE64_API_KEY = base64.b64encode(f"{STACKONE_API_KEY}:".encode()).decode()

# Create HTTP client with StackOne authentication headers
http_client = HttpxClientFactory(
    headers={
        "Authorization": f"Basic {BASE64_API_KEY}",
        "x-account-id": STACKONE_ACCOUNT_ID
    },
    timeout=120.0,  # StackOne agents run an LLM per call; the 5s default is too short
)

# Create a remote A2A agent pointing to StackOne
stackone_agent = A2aRemoteAgent(
    url="https://a2a.stackone.com",
    name="stackone_hr_agent",
    client=http_client
)

# Create a local agent to orchestrate conversations
llm_config = LLMConfig({"model": "gpt-4o-mini"})

orchestrator = ConversableAgent(
    name="hr_assistant",
    system_message="""
        You are an HR assistant that helps with employee management tasks.
        Delegate HR operations to the stackone_hr_agent.

        When the user asks about employees, time off, or other HR data,
        use the stackone_hr_agent to fetch or update the information.
    """,
    llm_config=llm_config,
)

# Start a conversation
async def main():
    await orchestrator.a_initiate_chat(
        recipient=stackone_agent,
        message={"role": "user", "content": "List the first 5 employees"},
        max_turns=1,
    )

asyncio.run(main())

See Authentication Guide for details on obtaining your API key and account ID.

Architecture

AG2’s A2aRemoteAgent allows your local agent to communicate with remote A2A agents:

Complete Example

Here’s a complete example with a local tool and a remote StackOne agent:

import asyncio
import base64
from datetime import datetime
from autogen import ConversableAgent, LLMConfig
from autogen.a2a import A2aRemoteAgent, HttpxClientFactory

# Configuration
STACKONE_API_KEY = "<stackone_api_key>"
STACKONE_ACCOUNT_ID = "<account_id>"
BASE64_API_KEY = base64.b64encode(f"{STACKONE_API_KEY}:".encode()).decode()

# Local tool example
def get_current_date() -> str:
    """Get the current date."""
    return datetime.now().strftime("%Y-%m-%d")

# Create HTTP client with StackOne authentication headers
http_client = HttpxClientFactory(
    headers={
        "Authorization": f"Basic {BASE64_API_KEY}",
        "x-account-id": STACKONE_ACCOUNT_ID
    },
    timeout=120.0,  # StackOne agents run an LLM per call; the 5s default is too short
)

# Remote StackOne A2A agent
stackone_agent = A2aRemoteAgent(
    url="https://a2a.stackone.com",
    name="stackone_hr_agent",
    client=http_client
)

# Configure local agent with LLM
llm_config = LLMConfig({"model": "gpt-4o-mini"})

# Root orchestrator agent
orchestrator = ConversableAgent(
    name="hr_assistant",
    system_message="""
        You are an HR assistant that helps with employee management tasks.

        You have access to:
        1. get_current_date - Get today's date
        2. stackone_hr_agent - Delegate HR operations (employees, time off, etc.)

        When users ask about HR data, delegate to stackone_hr_agent.
        Always confirm actions before making changes.
    """,
    llm_config=llm_config,
)

# Register local tool
orchestrator.register_for_llm(name="get_current_date", description="Get the current date")(get_current_date)

async def main():
    # Start conversation with StackOne agent
    await orchestrator.a_initiate_chat(
        recipient=stackone_agent,
        message={"role": "user", "content": "List all employees"},
        max_turns=1,
    )

    # Print the conversation history
    messages = orchestrator.chat_messages[stackone_agent.name]
    for message in messages:
        print(f"{message['name']}: {message['content']}")

asyncio.run(main())

Creating from AgentCard

If you already have an AgentCard (e.g., fetched from a discovery service), create an A2aRemoteAgent directly from it:

from a2a.types import AgentCard, AgentCapabilities
from autogen.a2a import A2aRemoteAgent

# Create or fetch an AgentCard
card = AgentCard(
    name="stackone_hr_agent",
    url="https://a2a.stackone.com",
    description="Agent that handles HR operations via StackOne integrations",
    version="0.1.0",
    default_input_modes=["text"],
    default_output_modes=["text"],
    capabilities=AgentCapabilities(streaming=True),
    skills=[],
    supports_authenticated_extended_card=False,
)

# Create the remote agent from the card
stackone_agent = A2aRemoteAgent.from_card(card)

Human in the Loop Support

A2aRemoteAgent automatically handles Human in the Loop (HITL) interactions when the remote agent requests human input. This happens transparently - the client loops until the agent completes its task or the conversation terminates.

from autogen.a2a import A2aRemoteAgent

# Connect to a remote agent that may request human input
stackone_agent = A2aRemoteAgent(
    url="https://a2a.stackone.com",
    name="approval_agent",
    client=http_client
)

# The client automatically handles any human input requests
await orchestrator.a_initiate_chat(
    recipient=stackone_agent,
    message={"role": "user", "content": "Should I approve this budget proposal?"}
)
# If the remote agent requests input, the user will be prompted automatically

Multiple StackOne Accounts

Connect to multiple StackOne accounts by creating separate HTTP clients and A2aRemoteAgent instances:

import base64
from autogen import ConversableAgent, LLMConfig
from autogen.a2a import A2aRemoteAgent, HttpxClientFactory

STACKONE_API_KEY = "<stackone_api_key>"
BASE64_API_KEY = base64.b64encode(f"{STACKONE_API_KEY}:".encode()).decode()

# HTTP client for HiBob account
hibob_http_client = HttpxClientFactory(
    headers={
        "Authorization": f"Basic {BASE64_API_KEY}",
        "x-account-id": "<hibob_account_id>"
    },
    timeout=120.0,
)

# HTTP client for BambooHR account
bamboo_http_client = HttpxClientFactory(
    headers={
        "Authorization": f"Basic {BASE64_API_KEY}",
        "x-account-id": "<bamboohr_account_id>"
    },
    timeout=120.0,
)

# Agent for HiBob integration
hibob_agent = A2aRemoteAgent(
    url="https://a2a.stackone.com",
    name="hibob_agent",
    client=hibob_http_client
)

# Agent for BambooHR integration
bamboo_agent = A2aRemoteAgent(
    url="https://a2a.stackone.com",
    name="bamboo_agent",
    client=bamboo_http_client
)

# Orchestrator with access to both agents
llm_config = LLMConfig({"model": "gpt-4o-mini"})

orchestrator = ConversableAgent(
    name="multi_hr_assistant",
    system_message="""
        You have access to multiple HR systems:
        - hibob_agent: For HiBob HR operations
        - bamboo_agent: For BambooHR operations

        Route requests to the appropriate system based on user context.
    """,
    llm_config=llm_config,
)

Advanced Configuration

Custom HTTP Client

Set custom options on the HTTP client such as headers and timeout:

from autogen.a2a import A2aRemoteAgent, HttpxClientFactory

# Create custom HTTP client factory
http_client = HttpxClientFactory(
    timeout=120.0,  # StackOne agents run an LLM per call; the 5s default is too short
    headers={
        "User-Agent": "MyApp/1.0",
        "Authorization": f"Basic {BASE64_API_KEY}",
        "x-account-id": STACKONE_ACCOUNT_ID
    }
)

stackone_agent = A2aRemoteAgent(
    url="https://a2a.stackone.com",
    name="stackone_hr_agent",
    client=http_client
)

Client Configuration

Pass additional client configuration using ClientConfig:

from a2a.client import ClientConfig
from autogen.a2a import A2aRemoteAgent

stackone_agent = A2aRemoteAgent(
    url="https://a2a.stackone.com",
    name="stackone_hr_agent",
    client=http_client,
    client_config=ClientConfig(streaming=True),
)

Testing with MockClient

Mock remote agent replies for testing:

from autogen.a2a import A2aRemoteAgent, MockClient

stackone_agent = A2aRemoteAgent(
    url="https://a2a.stackone.com",
    name="stackone_hr_agent",
    client=MockClient(response_message="Mock response for testing"),
)

Best Practices

Use Async Methods

A2aRemoteAgent only supports asynchronous methods. Always use a_initiate_chat and other async methods:

# Correct: Use async methods
await orchestrator.a_initiate_chat(
    recipient=stackone_agent,
    message={"role": "user", "content": "List employees"}
)

# Incorrect: Sync methods won't work
# orchestrator.initiate_chat(...)  # This will fail

Clear System Messages

Provide detailed system messages for your orchestrator agent. Include clear instructions about when to delegate to remote agents:

orchestrator = ConversableAgent(
    name="hr_assistant",
    system_message="""
        You are an HR assistant. When users ask about:
        - Employee data -> delegate to stackone_hr_agent
        - Time off requests -> delegate to stackone_hr_agent
        - Local operations -> use local tools

        Always confirm before making changes to HR systems.
    """,
    llm_config=llm_config,
)

Error Handling

Wrap A2A calls in try-except blocks to handle network or protocol errors:

async def safe_hr_query(message: str):
    try:
        await orchestrator.a_initiate_chat(
            recipient=stackone_agent,
            message={"role": "user", "content": message}
        )
    except Exception as e:
        print(f"Error communicating with HR agent: {e}")
        # Handle fallback logic

Message Format

Always provide messages in the correct format with role and content:

# Correct format
await orchestrator.a_initiate_chat(
    recipient=stackone_agent,
    message={"role": "user", "content": "List employees"}
)

# Also valid: just content string
await orchestrator.a_initiate_chat(
    recipient=stackone_agent,
    message="List employees"
)

Next Steps

A2A SDK

Build custom A2A tools with the Python or JavaScript SDK

Authentication

Learn about authentication and security

Quickstart

Learn A2A basics with cURL

Getting Started

Agent Frameworks

Agent Platforms

Overview

Installation

Quick Start

Architecture

Complete Example

Creating from AgentCard

Human in the Loop Support

Multiple StackOne Accounts

Advanced Configuration

Custom HTTP Client

Client Configuration

Testing with MockClient

Best Practices

Next Steps

A2A SDK

Authentication

Quickstart

​Overview

​Installation

​Quick Start

​Architecture

​Complete Example

​Creating from AgentCard

​Human in the Loop Support

​Multiple StackOne Accounts

​Advanced Configuration

​Custom HTTP Client

​Client Configuration

​Testing with MockClient

​Best Practices

​Next Steps

A2A SDK

Authentication

Quickstart

Overview

Installation

Quick Start

Architecture

Complete Example

Creating from AgentCard

Human in the Loop Support

Multiple StackOne Accounts

Advanced Configuration

Custom HTTP Client

Client Configuration

Testing with MockClient

Best Practices

Next Steps