Python: Add A2A server sample (#4528)

* Python: Add A2A server sample and fix client streaming bug

Add a pure Python A2A server sample so testing the A2A client no longer
requires running the .NET server. The server uses the a2a-sdk's
A2AStarletteApplication with uvicorn and supports three agent types
(invoice, policy, logistics) backed by AzureOpenAIResponsesClient.

New files:
- a2a_server.py: Main server entry point with CLI args
- agent_executor.py: Bridges a2a-sdk AgentExecutor to Agent Framework
- agent_definitions.py: Agent and AgentCard factory definitions
- invoice_data.py: Mock invoice data and query tool functions
- a2a_server.http: REST Client requests for testing

Also fixes a streaming bug in agent_with_a2a.py where async with was
used on ResponseStream which does not support the async context manager
protocol. Changed to async for to match all other samples.

Closes #4045

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address PR review: handle CancelledError and fix end_date filtering

- Re-raise asyncio.CancelledError before the broad exception handler
  so cooperative cancellation is not swallowed.
- Make end_date filter inclusive of the full day by comparing with
  < end + timedelta(days=1) instead of <= midnight.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This commit is contained in:
Giles Odigwe
2026-03-09 17:00:49 -07:00
committed by GitHub
Unverified
parent e2f0bc814e
commit ded32f3ff8
7 changed files with 709 additions and 25 deletions
+39 -16
View File
@@ -1,34 +1,57 @@
# A2A Agent Examples
This folder contains examples demonstrating how to create and use agents with the A2A (Agent2Agent) protocol from the `agent_framework` package to communicate with remote A2A agents.
This sample demonstrates how to host and consume agents using the [A2A (Agent2Agent) protocol](https://a2a-protocol.org/latest/) with the `agent_framework` package. There are two runnable entry points:
By default the A2AAgent waits for the remote agent to finish before returning (`background=False`), so long-running A2A tasks are handled transparently. For advanced scenarios where you need to poll or resubscribe to in-progress tasks using continuation tokens, see the [background responses sample](../../02-agents/background_responses.py).
| Run this file | To... |
|---------------|-------|
| **[`a2a_server.py`](a2a_server.py)** | Host an Agent Framework agent as an A2A-compliant server. |
| **[`agent_with_a2a.py`](agent_with_a2a.py)** | Connect to an A2A server and send requests (non-streaming and streaming). |
For more information about the A2A protocol specification, visit: https://a2a-protocol.org/latest/
## Examples
The remaining files are supporting modules used by the server:
| File | Description |
|------|-------------|
| [`agent_with_a2a.py`](agent_with_a2a.py) | Demonstrates agent discovery, non-streaming and streaming responses using the A2A protocol. |
| [`agent_definitions.py`](agent_definitions.py) | Agent and AgentCard factory definitions for invoice, policy, and logistics agents. |
| [`agent_executor.py`](agent_executor.py) | Bridges the a2a-sdk `AgentExecutor` interface to Agent Framework agents. |
| [`invoice_data.py`](invoice_data.py) | Mock invoice data and tool functions for the invoice agent. |
| [`a2a_server.http`](a2a_server.http) | REST Client requests for testing the server directly from VS Code. |
## Environment Variables
Make sure to set the following environment variables before running the example:
Make sure to set the following environment variables before running the examples:
### Required
- `A2A_AGENT_HOST`: URL of a single A2A agent (for simple sample, e.g., `http://localhost:5001/`)
### Required (Server)
- `AZURE_AI_PROJECT_ENDPOINT` — Your Azure AI Foundry project endpoint
- `AZURE_OPENAI_RESPONSES_DEPLOYMENT_NAME` — Model deployment name (e.g. `gpt-4o`)
### Required (Client)
- `A2A_AGENT_HOST` — URL of the A2A server (e.g. `http://localhost:5001/`)
## Quick Testing with .NET A2A Servers
## Quick Start
For quick testing and demonstration, you can use the pre-built .NET A2A servers from this repository:
All commands below should be run from this directory:
**Quick Testing Reference**: Use the .NET A2A Client Server sample at:
`..\agent-framework\dotnet\samples\05-end-to-end\A2AClientServer`
### Run Python A2A Sample
```powershell
# Simple A2A sample (single agent)
cd python/samples/04-hosting/a2a
```
### 1. Start the A2A Server
Pick an agent type and start the server (each in its own terminal):
```powershell
uv run python a2a_server.py --agent-type invoice --port 5000
uv run python a2a_server.py --agent-type policy --port 5001
uv run python a2a_server.py --agent-type logistics --port 5002
```
You can run one agent or all three — each listens on its own port.
### 2. Run the A2A Client
In a separate terminal (from the same directory), point the client at a running server:
```powershell
$env:A2A_AGENT_HOST = "http://localhost:5001/"
uv run python agent_with_a2a.py
```
@@ -0,0 +1,82 @@
### Each A2A agent is available at a different host address
@hostInvoice = http://localhost:5000
@hostPolicy = http://localhost:5001
@hostLogistics = http://localhost:5002
### Query agent card for the invoice agent
GET {{hostInvoice}}/.well-known/agent.json
### Send a message to the invoice agent
POST {{hostInvoice}}
Content-Type: application/json
{
"id": "1",
"jsonrpc": "2.0",
"method": "message/send",
"params": {
"message": {
"kind": "message",
"role": "user",
"messageId": "msg_1",
"parts": [
{
"kind": "text",
"text": "Show me all invoices for Contoso"
}
]
}
}
}
### Query agent card for the policy agent
GET {{hostPolicy}}/.well-known/agent.json
### Send a message to the policy agent
POST {{hostPolicy}}
Content-Type: application/json
{
"id": "2",
"jsonrpc": "2.0",
"method": "message/send",
"params": {
"message": {
"kind": "message",
"role": "user",
"messageId": "msg_2",
"parts": [
{
"kind": "text",
"text": "What is the policy for short shipments?"
}
]
}
}
}
### Query agent card for the logistics agent
GET {{hostLogistics}}/.well-known/agent.json
### Send a message to the logistics agent
POST {{hostLogistics}}
Content-Type: application/json
{
"id": "3",
"jsonrpc": "2.0",
"method": "message/send",
"params": {
"message": {
"kind": "message",
"role": "user",
"messageId": "msg_3",
"parts": [
{
"kind": "text",
"text": "What is the status for SHPMT-SAP-001?"
}
]
}
}
}
+120
View File
@@ -0,0 +1,120 @@
# Copyright (c) Microsoft. All rights reserved.
import argparse
import os
import sys
import uvicorn
from a2a.server.apps.jsonrpc.starlette_app import A2AStarletteApplication
from a2a.server.request_handlers.default_request_handler import DefaultRequestHandler
from a2a.server.tasks.inmemory_task_store import InMemoryTaskStore
from agent_definitions import AGENT_CARD_FACTORIES, AGENT_FACTORIES
from agent_executor import AgentFrameworkExecutor
from agent_framework.azure import AzureOpenAIResponsesClient
from azure.identity import AzureCliCredential
from dotenv import load_dotenv
# Load environment variables from .env file
load_dotenv()
"""
A2A Server Sample — Host an Agent Framework agent as an A2A endpoint
This sample creates a Python-based A2A-compliant server that wraps an Agent
Framework agent. The server uses the a2a-sdk's Starlette application to handle
JSON-RPC requests and serves the AgentCard at /.well-known/agent.json.
Three agent types are available:
- invoice — Answers invoice queries using mock data and function tools.
- policy — Returns a fixed policy response.
- logistics — Returns a fixed logistics response.
Usage:
uv run python a2a_server.py --agent-type policy --port 5001
uv run python a2a_server.py --agent-type invoice --port 5000
uv run python a2a_server.py --agent-type logistics --port 5002
Environment variables:
AZURE_AI_PROJECT_ENDPOINT — Your Azure AI Foundry project endpoint
AZURE_OPENAI_RESPONSES_DEPLOYMENT_NAME — Model deployment name (e.g. gpt-4o)
"""
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="A2A Agent Server")
parser.add_argument(
"--agent-type",
choices=["invoice", "policy", "logistics"],
default="policy",
help="Type of agent to host (default: policy)",
)
parser.add_argument(
"--host",
default="localhost",
help="Host to bind to (default: localhost)",
)
parser.add_argument(
"--port",
type=int,
default=5001,
help="Port to listen on (default: 5001)",
)
return parser.parse_args()
def main() -> None:
args = parse_args()
# Validate environment
project_endpoint = os.getenv("AZURE_AI_PROJECT_ENDPOINT")
deployment_name = os.getenv("AZURE_OPENAI_RESPONSES_DEPLOYMENT_NAME")
if not project_endpoint:
print("Error: AZURE_AI_PROJECT_ENDPOINT environment variable is not set.")
sys.exit(1)
if not deployment_name:
print("Error: AZURE_OPENAI_RESPONSES_DEPLOYMENT_NAME environment variable is not set.")
sys.exit(1)
# Create the LLM client
credential = AzureCliCredential()
client = AzureOpenAIResponsesClient(
project_endpoint=project_endpoint,
deployment_name=deployment_name,
credential=credential,
)
# Create the Agent Framework agent for the chosen type
agent_factory = AGENT_FACTORIES[args.agent_type]
agent = agent_factory(client)
# Build the A2A server components
url = f"http://{args.host}:{args.port}/"
agent_card = AGENT_CARD_FACTORIES[args.agent_type](url)
executor = AgentFrameworkExecutor(agent)
task_store = InMemoryTaskStore()
request_handler = DefaultRequestHandler(
agent_executor=executor,
task_store=task_store,
)
a2a_app = A2AStarletteApplication(
agent_card=agent_card,
http_handler=request_handler,
)
print(f"Starting A2A server: {agent_card.name}")
print(f" Agent type : {args.agent_type}")
print(f" Listening : {url}")
print(f" Agent card : {url}.well-known/agent.json")
print()
uvicorn.run(
a2a_app.build(),
host=args.host,
port=args.port,
)
if __name__ == "__main__":
main()
@@ -0,0 +1,169 @@
# Copyright (c) Microsoft. All rights reserved.
"""Agent definitions and AgentCard factories for the A2A server sample.
Provides factory functions to create Agent Framework agents and A2A
AgentCards for the invoice, policy, and logistics agent types.
"""
from __future__ import annotations
from typing import TYPE_CHECKING
from a2a.types import AgentCapabilities, AgentCard, AgentSkill
from invoice_data import query_by_invoice_id, query_by_transaction_id, query_invoices
if TYPE_CHECKING:
from agent_framework import Agent
from agent_framework.azure import AzureOpenAIResponsesClient
# ---------------------------------------------------------------------------
# Agent instructions
# ---------------------------------------------------------------------------
INVOICE_INSTRUCTIONS = "You specialize in handling queries related to invoices."
POLICY_INSTRUCTIONS = """\
You specialize in handling queries related to policies and customer communications.
Always reply with exactly this text:
Policy: Short Shipment Dispute Handling Policy V2.1
Summary: "For short shipments reported by customers, first verify internal shipment records
(SAP) and physical logistics scan data (BigQuery). If discrepancy is confirmed and logistics data
shows fewer items packed than invoiced, issue a credit for the missing items. Document the
resolution in SAP CRM and notify the customer via email within 2 business days, referencing the
original invoice and the credit memo number. Use the 'Formal Credit Notification' email
template."
"""
LOGISTICS_INSTRUCTIONS = """\
You specialize in handling queries related to logistics.
Always reply with exactly:
Shipment number: SHPMT-SAP-001
Item: TSHIRT-RED-L
Quantity: 900
"""
# ---------------------------------------------------------------------------
# Agent factories
# ---------------------------------------------------------------------------
def create_invoice_agent(client: AzureOpenAIResponsesClient) -> Agent:
"""Create an invoice agent backed by the given client with query tools."""
return client.as_agent(
name="InvoiceAgent",
instructions=INVOICE_INSTRUCTIONS,
tools=[query_invoices, query_by_transaction_id, query_by_invoice_id],
)
def create_policy_agent(client: AzureOpenAIResponsesClient) -> Agent:
"""Create a policy agent backed by the given client."""
return client.as_agent(
name="PolicyAgent",
instructions=POLICY_INSTRUCTIONS,
)
def create_logistics_agent(client: AzureOpenAIResponsesClient) -> Agent:
"""Create a logistics agent backed by the given client."""
return client.as_agent(
name="LogisticsAgent",
instructions=LOGISTICS_INSTRUCTIONS,
)
# ---------------------------------------------------------------------------
# AgentCard factories
# ---------------------------------------------------------------------------
_CAPABILITIES = AgentCapabilities(streaming=True, push_notifications=False)
def get_invoice_agent_card(url: str) -> AgentCard:
"""Return an A2A AgentCard for the invoice agent."""
return AgentCard(
name="InvoiceAgent",
description="Handles requests relating to invoices.",
url=url,
version="1.0.0",
default_input_modes=["text"],
default_output_modes=["text"],
capabilities=_CAPABILITIES,
skills=[
AgentSkill(
id="id_invoice_agent",
name="InvoiceQuery",
description="Handles requests relating to invoices.",
tags=["invoice", "agent-framework"],
examples=["List the latest invoices for Contoso."],
),
],
)
def get_policy_agent_card(url: str) -> AgentCard:
"""Return an A2A AgentCard for the policy agent."""
return AgentCard(
name="PolicyAgent",
description="Handles requests relating to policies and customer communications.",
url=url,
version="1.0.0",
default_input_modes=["text"],
default_output_modes=["text"],
capabilities=_CAPABILITIES,
skills=[
AgentSkill(
id="id_policy_agent",
name="PolicyAgent",
description="Handles requests relating to policies and customer communications.",
tags=["policy", "agent-framework"],
examples=["What is the policy for short shipments?"],
),
],
)
def get_logistics_agent_card(url: str) -> AgentCard:
"""Return an A2A AgentCard for the logistics agent."""
return AgentCard(
name="LogisticsAgent",
description="Handles requests relating to logistics.",
url=url,
version="1.0.0",
default_input_modes=["text"],
default_output_modes=["text"],
capabilities=_CAPABILITIES,
skills=[
AgentSkill(
id="id_logistics_agent",
name="LogisticsQuery",
description="Handles requests relating to logistics.",
tags=["logistics", "agent-framework"],
examples=["What is the status for SHPMT-SAP-001"],
),
],
)
# ---------------------------------------------------------------------------
# Lookup helpers
# ---------------------------------------------------------------------------
AGENT_FACTORIES = {
"invoice": create_invoice_agent,
"policy": create_policy_agent,
"logistics": create_logistics_agent,
}
AGENT_CARD_FACTORIES = {
"invoice": get_invoice_agent_card,
"policy": get_policy_agent_card,
"logistics": get_logistics_agent_card,
}
@@ -0,0 +1,123 @@
# Copyright (c) Microsoft. All rights reserved.
"""AgentExecutor bridge between the a2a-sdk server and Agent Framework agents.
Implements the a2a-sdk ``AgentExecutor`` interface so that incoming A2A
requests are forwarded to an Agent Framework agent and the response is
published back through the a2a-sdk event queue.
"""
from __future__ import annotations
import asyncio
import uuid
from typing import TYPE_CHECKING
from a2a.server.agent_execution.agent_executor import AgentExecutor
from a2a.types import (
Message,
Part,
Role,
TaskState,
TaskStatus,
TaskStatusUpdateEvent,
TextPart,
)
if TYPE_CHECKING:
from a2a.server.agent_execution.context import RequestContext
from a2a.server.events.event_queue import EventQueue
from agent_framework import Agent
class AgentFrameworkExecutor(AgentExecutor):
"""Bridges A2A protocol requests to an Agent Framework agent.
For each incoming ``execute`` call the executor:
1. Extracts the user's text from the A2A ``RequestContext``.
2. Runs the Agent Framework agent (non-streaming).
3. Publishes the result as an A2A ``Message`` to the ``EventQueue``.
"""
def __init__(self, agent: Agent) -> None:
self.agent = agent
async def execute(self, context: RequestContext, event_queue: EventQueue) -> None:
"""Run the agent and publish the response."""
user_text = context.get_user_input()
if not user_text:
user_text = "Hello"
task_id = context.task_id or str(uuid.uuid4())
context_id = context.context_id or str(uuid.uuid4())
# Signal that the agent is working
await event_queue.enqueue_event(
TaskStatusUpdateEvent(
task_id=task_id,
context_id=context_id,
status=TaskStatus(state=TaskState.working),
final=False,
)
)
try:
response = await self.agent.run(user_text)
# Build response text from agent messages
response_parts: list[Part] = []
for msg in response.messages:
if msg.text:
response_parts.append(TextPart(text=msg.text))
if not response_parts:
response_parts.append(TextPart(text=str(response)))
# Publish the agent's response as a completed message
await event_queue.enqueue_event(
TaskStatusUpdateEvent(
task_id=task_id,
context_id=context_id,
status=TaskStatus(
state=TaskState.completed,
message=Message(
message_id=str(uuid.uuid4()),
role=Role.agent,
parts=response_parts,
),
),
final=True,
)
)
except asyncio.CancelledError:
raise
except Exception as e:
await event_queue.enqueue_event(
TaskStatusUpdateEvent(
task_id=task_id,
context_id=context_id,
status=TaskStatus(
state=TaskState.failed,
message=Message(
message_id=str(uuid.uuid4()),
role=Role.agent,
parts=[TextPart(text=f"Agent error: {e}")],
),
),
final=True,
)
)
async def cancel(self, context: RequestContext, event_queue: EventQueue) -> None:
"""Handle cancellation by publishing a canceled status."""
task_id = context.task_id or str(uuid.uuid4())
context_id = context.context_id or str(uuid.uuid4())
await event_queue.enqueue_event(
TaskStatusUpdateEvent(
task_id=task_id,
context_id=context_id,
status=TaskStatus(state=TaskState.canceled),
final=True,
)
)
@@ -78,16 +78,16 @@ async def main():
# Updates arrive as Server-Sent Events, letting you observe
# progress in real time as the remote agent works.
print("\n--- Streaming response ---")
async with agent.run("Tell me about yourself", stream=True) as stream:
async for update in stream:
for content in update.contents:
if content.text:
print(f" {content.text}")
stream = agent.run("Tell me about yourself", stream=True)
async for update in stream:
for content in update.contents:
if content.text:
print(f" {content.text}")
response = await stream.get_final_response()
print(f"\nFinal response ({len(response.messages)} message(s)):")
for message in response.messages:
print(f" {message.text}")
response = await stream.get_final_response()
print(f"\nFinal response ({len(response.messages)} message(s)):")
for message in response.messages:
print(f" {message.text}")
if __name__ == "__main__":
@@ -0,0 +1,167 @@
# Copyright (c) Microsoft. All rights reserved.
"""Mock invoice data and tool functions for the A2A server sample.
Provides mock invoice data and query tools for the A2A server sample,
enabling invoice-related queries through the A2A protocol.
"""
import json
import random
from dataclasses import dataclass, field
from datetime import datetime, timedelta, timezone
from typing import Annotated
from agent_framework import tool
from pydantic import Field
@dataclass
class Product:
"""A product line item on an invoice."""
name: str
quantity: int
price_per_unit: float
@property
def total_price(self) -> float:
return self.quantity * self.price_per_unit
def to_dict(self) -> dict:
return {
"name": self.name,
"quantity": self.quantity,
"price_per_unit": self.price_per_unit,
"total_price": self.total_price,
}
@dataclass
class Invoice:
"""An invoice record with products."""
transaction_id: str
invoice_id: str
company_name: str
invoice_date: datetime
products: list[Product] = field(default_factory=list)
@property
def total_invoice_price(self) -> float:
return sum(p.total_price for p in self.products)
def to_dict(self) -> dict:
return {
"transaction_id": self.transaction_id,
"invoice_id": self.invoice_id,
"company_name": self.company_name,
"invoice_date": self.invoice_date.strftime("%Y-%m-%d"),
"products": [p.to_dict() for p in self.products],
"total_invoice_price": self.total_invoice_price,
}
def _random_date_within_last_two_months() -> datetime:
end_date = datetime.now(timezone.utc)
start_date = end_date - timedelta(days=60)
random_days = random.randint(0, 60)
return start_date + timedelta(days=random_days)
def _build_invoices() -> list[Invoice]:
"""Build 10 mock invoices."""
return [
Invoice("TICKET-XYZ987", "INV789", "Contoso", _random_date_within_last_two_months(), [
Product("T-Shirts", 150, 10.00),
Product("Hats", 200, 15.00),
Product("Glasses", 300, 5.00),
]),
Invoice("TICKET-XYZ111", "INV111", "XStore", _random_date_within_last_two_months(), [
Product("T-Shirts", 2500, 12.00),
Product("Hats", 1500, 8.00),
Product("Glasses", 200, 20.00),
]),
Invoice("TICKET-XYZ222", "INV222", "Cymbal Direct", _random_date_within_last_two_months(), [
Product("T-Shirts", 1200, 14.00),
Product("Hats", 800, 7.00),
Product("Glasses", 500, 25.00),
]),
Invoice("TICKET-XYZ333", "INV333", "Contoso", _random_date_within_last_two_months(), [
Product("T-Shirts", 400, 11.00),
Product("Hats", 600, 15.00),
Product("Glasses", 700, 5.00),
]),
Invoice("TICKET-XYZ444", "INV444", "XStore", _random_date_within_last_two_months(), [
Product("T-Shirts", 800, 10.00),
Product("Hats", 500, 18.00),
Product("Glasses", 300, 22.00),
]),
Invoice("TICKET-XYZ555", "INV555", "Cymbal Direct", _random_date_within_last_two_months(), [
Product("T-Shirts", 1100, 9.00),
Product("Hats", 900, 12.00),
Product("Glasses", 1200, 15.00),
]),
Invoice("TICKET-XYZ666", "INV666", "Contoso", _random_date_within_last_two_months(), [
Product("T-Shirts", 2500, 8.00),
Product("Hats", 1200, 10.00),
Product("Glasses", 1000, 6.00),
]),
Invoice("TICKET-XYZ777", "INV777", "XStore", _random_date_within_last_two_months(), [
Product("T-Shirts", 1900, 13.00),
Product("Hats", 1300, 16.00),
Product("Glasses", 800, 19.00),
]),
Invoice("TICKET-XYZ888", "INV888", "Cymbal Direct", _random_date_within_last_two_months(), [
Product("T-Shirts", 2200, 11.00),
Product("Hats", 1700, 8.50),
Product("Glasses", 600, 21.00),
]),
Invoice("TICKET-XYZ999", "INV999", "Contoso", _random_date_within_last_two_months(), [
Product("T-Shirts", 1400, 10.50),
Product("Hats", 1100, 9.00),
Product("Glasses", 950, 12.00),
]),
]
# Module-level singleton so dates are stable for the lifetime of the server
INVOICES = _build_invoices()
@tool(approval_mode="never_require")
def query_invoices(
company_name: Annotated[str, Field(description="The company name to filter invoices by.")],
start_date: Annotated[str | None, Field(description="Optional start date (YYYY-MM-DD) to filter invoices.")] = None,
end_date: Annotated[str | None, Field(description="Optional end date (YYYY-MM-DD) to filter invoices.")] = None,
) -> str:
"""Retrieves invoices for the specified company and optionally within the specified time range."""
results = [i for i in INVOICES if i.company_name.lower() == company_name.lower()]
if start_date:
start = datetime.strptime(start_date, "%Y-%m-%d").replace(tzinfo=timezone.utc)
results = [i for i in results if i.invoice_date >= start]
if end_date:
end = datetime.strptime(end_date, "%Y-%m-%d").replace(tzinfo=timezone.utc) + timedelta(days=1)
results = [i for i in results if i.invoice_date < end]
return json.dumps([i.to_dict() for i in results], indent=2)
@tool(approval_mode="never_require")
def query_by_transaction_id(
transaction_id: Annotated[str, Field(description="The transaction ID to look up (e.g. TICKET-XYZ987).")],
) -> str:
"""Retrieves invoice using the transaction id."""
results = [i for i in INVOICES if i.transaction_id.lower() == transaction_id.lower()]
return json.dumps([i.to_dict() for i in results], indent=2)
@tool(approval_mode="never_require")
def query_by_invoice_id(
invoice_id: Annotated[str, Field(description="The invoice ID to look up (e.g. INV789).")],
) -> str:
"""Retrieves invoice using the invoice id."""
results = [i for i in INVOICES if i.invoice_id.lower() == invoice_id.lower()]
return json.dumps([i.to_dict() for i in results], indent=2)