Overview The Microsoft Agent Framework provides multiple hosting options for deploying agents to production environments. Whether you need stateless HTTP endpoints, durable orchestrations, or agent-to-agent communication, the framework offers flexible deployment patterns.
Hosting Options Azure Functions Host agents as HTTP-accessible services using Azure Functions with built-in scaling, monitoring, and Azure integration.
Best for:
Stateless request/response patterns
HTTP API endpoints
Serverless deployments
Azure ecosystem integration
Key Features:
Auto-generated HTTP endpoints per agent
Built-in health checks
Session management
Multi-agent hosting
Learn more about Azure Functions hosting → DurableTask Integration Leverage Durable Functions and DurableTask for long-running workflows, orchestrations, and stateful agent execution.
Best for:
Long-running workflows
Multi-agent orchestrations
Human-in-the-loop patterns
Reliable state management
Key Features:
Durable orchestrations
Checkpoint and resume
Parallel agent execution
External event handling
Learn more about DurableTask → A2A Protocol Connect agents across systems using the Agent-to-Agent (A2A) protocol for standardized inter-agent communication.
Best for:
Distributed agent systems
Cross-framework communication
Agent discovery
Standardized interfaces
Key Features:
Standard protocol compliance
Agent discovery via AgentCard
Streaming and non-streaming modes
Background task handling
Learn more about A2A Protocol → Deployment Patterns Single Agent Hosting Deploy a single agent with dedicated endpoints:
from agent_framework.azure import AgentFunctionApp, AzureOpenAIChatClient from azure.identity import AzureCliCredential agent = AzureOpenAIChatClient( credential = AzureCliCredential()).as_agent( name = "Assistant" , instructions = "You are a helpful assistant." ) app = AgentFunctionApp( agents = [agent], enable_health_check = True )
Multi-Agent Hosting Host multiple specialized agents in a single application:
weather_agent = client.as_agent( name = "WeatherAgent" , instructions = "Provide weather information." , tools = [get_weather] ) math_agent = client.as_agent( name = "MathAgent" , instructions = "Help with calculations." , tools = [calculate_tip] ) app = AgentFunctionApp( agents = [weather_agent, math_agent])
Orchestrated Workflows Create complex workflows with multiple agents:
@app.orchestration_trigger ( context_name = "context" ) def agent_workflow ( context : DurableOrchestrationContext): """Sequential agent execution with shared session.""" agent = app.get_agent(context, "WriterAgent" ) session = agent.create_session() # First pass initial = yield agent.run( messages = "Write a sentence about AI." , session = session ) # Second pass with context refined = yield agent.run( messages = f "Improve this: { initial.text } " , session = session ) return refined.text
Configuration Environment Variables All hosting patterns require Azure OpenAI configuration:
export AZURE_OPENAI_ENDPOINT = "https://your-resource.openai.azure.com/" export AZURE_OPENAI_DEPLOYMENT_NAME = "gpt-4o-mini" # Optional: Use API key instead of Azure CLI authentication export AZURE_OPENAI_API_KEY = "your-api-key"
Azure Functions Configuration For Azure Functions, configure local.settings.json:
{ "IsEncrypted" : false , "Values" : { "FUNCTIONS_WORKER_RUNTIME" : "python" , "AzureWebJobsStorage" : "UseDevelopmentStorage=true" , "DURABLE_TASK_SCHEDULER_CONNECTION_STRING" : "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None" , "TASKHUB_NAME" : "default" , "AZURE_OPENAI_ENDPOINT" : "<AZURE_OPENAI_ENDPOINT>" , "AZURE_OPENAI_CHAT_DEPLOYMENT_NAME" : "<AZURE_OPENAI_CHAT_DEPLOYMENT_NAME>" } }
Prerequisites
Azure Functions Core Tools : Install v4.x Azure CLI : For authentication (az login)Docker : For local emulators (Azurite, DTS)
Azure Services
Azure OpenAI Service : Deploy a chat model (gpt-4o-mini or better)Azure Storage : Required for Azure Functions stateDurable Task Scheduler (optional): For orchestrations
Local Emulators Start Azurite for Azure Storage emulation:
docker run -d --name storage-emulator -p 10000:10000 -p 10001:10001 -p 10002:10002 mcr.microsoft.com/azure-storage/azurite
Start Durable Task Scheduler emulator:
docker run -d --name dts-emulator -p 8080:8080 -p 8082:8082 mcr.microsoft.com/dts/dts-emulator:latest
The DTS dashboard will be available at http://localhost:8082.
Authentication The framework supports multiple authentication methods:
Azure CLI (Development) from azure.identity import AzureCliCredential credential = AzureCliCredential() client = AzureOpenAIChatClient( credential = credential)
API Key (Testing) import os client = AzureOpenAIChatClient( api_key = os.environ[ "AZURE_OPENAI_API_KEY" ] )
Managed Identity (Production) from azure.identity import ManagedIdentityCredential credential = ManagedIdentityCredential() client = AzureOpenAIChatClient( credential = credential)
Production Considerations : DefaultAzureCredential is convenient for development but may cause latency in production due to credential probing. Use specific credentials like ManagedIdentityCredential for production deployments.
Testing Deployed Agents Test HTTP endpoints using curl, REST clients, or HTTP files:
# Test single agent curl -X POST http://localhost:7071/api/agents/Assistant/run \ -H "Content-Type: text/plain" \ -d "What is machine learning?" # Test with JSON curl -X POST http://localhost:7071/api/agents/Assistant/run \ -H "Content-Type: application/json" \ -d '{"message": "Explain AI", "session_id": "user-123"}' # Test orchestration curl -X POST http://localhost:7071/api/orchestration/run
Monitoring and Observability Azure Application Insights Enable Application Insights for production monitoring:
{ "Values" : { "APPLICATIONINSIGHTS_CONNECTION_STRING" : "InstrumentationKey=..." } }
Logging Configure logging levels in host.json:
{ "version" : "2.0" , "logging" : { "logLevel" : { "default" : "Information" , "Microsoft.Agents" : "Debug" } } }
DTS Dashboard Monitor orchestrations at http://localhost:8082 (local) or your DTS endpoint (Azure).
Next Steps
Azure Functions Deploy agents as serverless functions
DurableTask Build long-running workflows
A2A Protocol Connect distributed agents