![\"Screenshot](\"https:\/\/devblogs.microsoft.com\/semantic-kernel\/wp-content\/uploads\/sites\/78\/2025\/07\/Screenshot-2025-07-18-at-8.48.46\u202fPM-300x105.png\")
<\/a><\/p>\nDevelopment Environment Setup<\/h3>\n
To build this A2A-enabled Semantic Kernel system, you’ll need:<\/p>\n
# Initialize Python project with uv (recommended)\r\nuv init multi_agent_system\r\ncd multi_agent_system\r\n\r\n# Core dependencies for Semantic Kernel and Azure integration\r\nuv add semantic-kernel[azure]\r\nuv add azure-identity\r\nuv add azure-ai-agents\r\nuv add python-dotenv\r\n\r\n# A2A protocol dependencies\r\nuv add a2a-client\r\nuv add httpx\r\n\r\n# MCP integration dependencies\r\nuv add semantic-kernel[mcp]\r\n\r\n# Web interface dependencies\r\nuv add gradio\r\n\r\n# Development dependencies\r\nuv add --dev pytest pytest-asyncio<\/code><\/pre>\nEnvironment Configuration<\/h3>\n
Configure your environment variables:<\/p>\n
# Azure AI Foundry configuration\r\nAZURE_AI_AGENT_ENDPOINT=https:\/\/your-ai-foundry-endpoint.azure.com\r\nAZURE_AI_AGENT_MODEL_DEPLOYMENT_NAME=Your AI Foundry Model Deployment Name\r\n\r\n# Remote agent endpoints\r\nPLAYWRIGHT_AGENT_URL=http:\/\/localhost:10001\r\nTOOL_AGENT_URL=http:\/\/localhost:10002\r\n\r\n# Optional: MCP server configuration\r\nMCP_SSE_URL=http:\/\/localhost:7071\/runtime\/webhooks\/mcp\/sse<\/code><\/pre>\nCreating the Central Routing Agent<\/h3>\n
The heart of our system is the routing agent that uses Azure AI Foundry for intelligent task delegation:<\/p>\n
import json\r\nimport os\r\nimport time\r\nimport uuid\r\nfrom typing import Any, Dict, List\r\nimport httpx\r\nfrom a2a.client import A2ACardResolver\r\nfrom azure.ai.agents import AgentsClient\r\nfrom azure.identity import DefaultAzureCredential\r\nfrom dotenv import load_dotenv\r\n\r\nclass RoutingAgent:\r\n \"\"\"Central routing agent powered by Azure AI Foundry.\"\"\"\r\n \r\n def __init__(self):\r\n self.remote_agent_connections = {}\r\n self.cards = {}\r\n self.agents_client = AgentsClient(\r\n endpoint=os.environ[\"AZURE_AI_AGENT_ENDPOINT\"],\r\n credential=DefaultAzureCredential(),\r\n )\r\n self.azure_agent = None\r\n self.current_thread = None\r\n\r\n async def initialize(self, remote_agent_addresses: list[str]):\r\n \"\"\"Initialize with A2A agent discovery.\"\"\"\r\n # Discover remote agents via A2A protocol\r\n async with httpx.AsyncClient(timeout=30) as client:\r\n for address in remote_agent_addresses:\r\n try:\r\n card_resolver = A2ACardResolver(client, address)\r\n card = await card_resolver.get_agent_card()\r\n \r\n from remote_agent_connection import RemoteAgentConnections\r\n remote_connection = RemoteAgentConnections(\r\n agent_card=card, agent_url=address\r\n )\r\n self.remote_agent_connections[card.name] = remote_connection\r\n self.cards[card.name] = card\r\n \r\n except Exception as e:\r\n print(f'Failed to connect to agent at {address}: {e}')\r\n\r\n # Create Azure AI agent for intelligent routing\r\n await self._create_azure_agent()\r\n\r\n async def _create_azure_agent(self):\r\n \"\"\"Create Azure AI agent with function calling capabilities.\"\"\"\r\n instructions = self._get_routing_instructions()\r\n \r\n # Define function for task delegation\r\n tools = [{\r\n \"type\": \"function\",\r\n \"function\": {\r\n \"name\": \"send_message\",\r\n \"description\": \"Delegate task to specialized remote agent\",\r\n \"parameters\": {\r\n \"type\": \"object\",\r\n \"properties\": {\r\n \"agent_name\": {\"type\": \"string\"},\r\n \"task\": {\"type\": \"string\"}\r\n },\r\n \"required\": [\"agent_name\", \"task\"]\r\n }\r\n }\r\n }]\r\n\r\n model_name = os.environ.get(\"AZURE_AI_AGENT_MODEL_DEPLOYMENT_NAME\", \"gpt-4\")\r\n self.azure_agent = self.agents_client.create_agent(\r\n model=model_name,\r\n name=\"routing-agent\",\r\n instructions=instructions,\r\n tools=tools\r\n )\r\n \r\n self.current_thread = self.agents_client.threads.create()\r\n print(f\"Routing agent initialized: {self.azure_agent.id}\")\r\n\r\n def _get_routing_instructions(self) -> str:\r\n \"\"\"Generate context-aware routing instructions.\"\"\"\r\n agent_info = [\r\n {'name': card.name, 'description': card.description}\r\n for card in self.cards.values()\r\n ]\r\n \r\n return f\"\"\"You are an intelligent routing agent for a multi-agent system.\r\n\r\nAvailable Specialist Agents:\r\n{json.dumps(agent_info, indent=2)}<\/code><\/pre>\n\n
Routing Guidelines:<\/div>\n
\n- Web automation, screenshots, browser tasks \u2192 Playwright Agent<\/li>\n
- Development tasks, file operations, repository management \u2192 Tool Agent<\/li>\n
- Always provide comprehensive task context when delegating<\/li>\n
- Explain to users what specialist is handling their request<\/li>\n<\/ul>\n<\/div>\n
Building Specialized Agents with MCP Integration<\/h3>\n
Our remote agents leverage Semantic Kernel’s MCP integration for extensible capabilities:<\/p>\n
from semantic_kernel.agents import AzureAIAgent, AzureAIAgentSettings\r\nfrom semantic_kernel.connectors.mcp import MCPStdioPlugin, MCPSsePlugin\r\nfrom azure.identity.aio import DefaultAzureCredential\r\n\r\nclass SemanticKernelMCPAgent:\r\n \"\"\"Specialized agent with MCP plugin integration.\"\"\"\r\n \r\n def __init__(self):\r\n self.agent = None\r\n self.client = None\r\n self.credential = None\r\n self.plugins = []\r\n\r\n async def initialize_playwright_agent(self):\r\n \"\"\"Initialize with Playwright automation via MCP STDIO.\"\"\"\r\n try:\r\n self.credential = DefaultAzureCredential()\r\n self.client = await AzureAIAgent.create_client(\r\n credential=self.credential\r\n ).__aenter__()\r\n \r\n # Create Playwright MCP plugin\r\n playwright_plugin = MCPStdioPlugin(\r\n name=\"Playwright\",\r\n command=\"npx\",\r\n args=[\"@playwright\/mcp@latest\"],\r\n )\r\n \r\n await playwright_plugin.__aenter__()\r\n self.plugins.append(playwright_plugin)\r\n \r\n # Create specialized agent\r\n agent_definition = await self.client.agents.create_agent(\r\n model=AzureAIAgentSettings().model_deployment_name,\r\n name=\"PlaywrightAgent\",\r\n instructions=(\r\n \"You are a web automation specialist. Use Playwright to \"\r\n \"navigate websites, take screenshots, interact with elements, \"\r\n \"and perform browser automation tasks.\"\r\n ),\r\n )\r\n\r\n self.agent = AzureAIAgent(\r\n client=self.client,\r\n definition=agent_definition,\r\n plugins=self.plugins,\r\n )\r\n \r\n except Exception as e:\r\n await self.cleanup()\r\n raise\r\n\r\n async def initialize_tools_agent(self, mcp_url: str):\r\n \"\"\"Initialize with development tools via MCP SSE.\"\"\"\r\n try:\r\n self.credential = DefaultAzureCredential()\r\n self.client = AzureAIAgent.create_client(credential=self.credential)\r\n \r\n # Create development tools MCP plugin\r\n tools_plugin = MCPSsePlugin(\r\n name=\"DevTools\",\r\n url=mcp_url,\r\n )\r\n \r\n await tools_plugin.__aenter__()\r\n self.plugins.append(tools_plugin)\r\n \r\n agent_definition = await self.client.agents.create_agent(\r\n model=AzureAIAgentSettings().model_deployment_name,\r\n name=\"DevAssistant\",\r\n instructions=(\r\n \"You are a development assistant. Help with repository \"\r\n \"management, file operations, opening projects in VS Code, \"\r\n \"and other development tasks.\"\r\n ),\r\n )\r\n\r\n self.agent = AzureAIAgent(\r\n client=self.client,\r\n definition=agent_definition,\r\n plugins=self.plugins,\r\n )\r\n \r\n except Exception as e:\r\n await self.cleanup()\r\n raise\r\n\r\n async def invoke(self, user_input: str) -> dict[str, Any]:\r\n \"\"\"Process tasks through the specialized agent.\"\"\"\r\n if not self.agent:\r\n return {\r\n 'is_task_complete': False,\r\n 'content': 'Agent not initialized.',\r\n }\r\n\r\n try:\r\n responses = []\r\n async for response in self.agent.invoke(\r\n messages=user_input,\r\n thread=self.thread,\r\n ):\r\n responses.append(str(response))\r\n self.thread = response.thread\r\n\r\n return {\r\n 'is_task_complete': True,\r\n 'content': \"\\n\".join(responses) or \"No response received.\",\r\n }\r\n except Exception as e:\r\n return {\r\n 'is_task_complete': False,\r\n 'content': f'Error: {str(e)}',\r\n }<\/code><\/pre>\nWeb Interface with Gradio<\/h3>\n
The system provides a modern chat interface powered by Gradio:<\/p>\n
import asyncio\r\nimport gradio as gr\r\nfrom routing_agent import RoutingAgent\r\n\r\nasync def get_response_from_agent(\r\n message: str, history: list[gr.ChatMessage]\r\n) -> gr.ChatMessage:\r\n \"\"\"Process user messages through the routing system.\"\"\"\r\n global ROUTING_AGENT\r\n \r\n try:\r\n response = await ROUTING_AGENT.process_user_message(message)\r\n return gr.ChatMessage(role=\"assistant\", content=response)\r\n except Exception as e:\r\n return gr.ChatMessage(\r\n role=\"assistant\",\r\n content=f\"\u274c Error: {str(e)}\"\r\n )\r\n\r\nasync def main():\r\n \"\"\"Launch the multi-agent system.\"\"\"\r\n # Initialize routing agent\r\n global ROUTING_AGENT\r\n ROUTING_AGENT = await RoutingAgent.create([\r\n os.getenv('PLAYWRIGHT_AGENT_URL', 'http:\/\/localhost:10001'),\r\n os.getenv('TOOL_AGENT_URL', 'http:\/\/localhost:10002'),\r\n ])\r\n ROUTING_AGENT.create_agent()\r\n\r\n # Create Gradio interface\r\n with gr.Blocks(theme=gr.themes.Ocean()) as demo:\r\n gr.Markdown(\"# \ud83e\udd16 Azure AI Multi-Agent System\")\r\n \r\n gr.ChatInterface(\r\n get_response_from_agent,\r\n title=\"Chat with AI Agents\",\r\n examples=[\r\n \"Navigate to github.com\/microsoft and take a screenshot\",\r\n \"Clone repository https:\/\/github.com\/microsoft\/semantic-kernel\",\r\n \"Open the cloned project in VS Code\",\r\n ]\r\n )\r\n\r\n demo.launch(server_name=\"0.0.0.0\", server_port=8083)<\/code><\/pre>\nDeployment and Operations<\/h3>\n
Local Development:<\/strong><\/p>\n# Start MCP server (Azure Functions)\r\ncd mcp_sse_server\/MCPAzureFunc\r\nfunc start\r\n\r\n# Start remote agents in separate terminals\r\ncd remote_agents\/playwright_agent && uv run .\r\ncd remote_agents\/tool_agent && uv run .\r\n\r\n# Start the host agent with web interface\r\ncd host_agent && uv run .<\/code><\/pre>\nProduction Considerations:<\/strong><\/p>\n
![\"arch](\"https:\/\/devblogs.microsoft.com\/semantic-kernel\/wp-content\/uploads\/sites\/78\/2025\/07\/arch.png\")
<\/a><\/p>\n