This directory contains samples for durable agent hosting using the Durable Task Scheduler. These samples demonstrate the worker-client architecture pattern, enabling distributed agent execution with persistent conversation state.

These samples import the durable hosting types directly from the extension packages — agent_framework_durabletask and agent_framework_azurefunctions:

from agent_framework_durabletask import DurableAIAgentWorker, DurableWorkflowClient
from agent_framework_azurefunctions import AgentFunctionApp

For backward compatibility these entry-point types are also re-exported from agent_framework.azure in the core agent-framework package, so existing from agent_framework.azure import ... code keeps working. New and updated samples should use the direct package imports shown above — the canonical, self-contained path for this repo — rather than routing through the agent_framework.azure shim.

Install and verify these tools before Running the Samples:

• Docker – run the Durable Task Scheduler emulator locally
• uv – manage Python dependencies (optional but recommended)
• Azure CLI – authenticate with az login for AzureCliCredential

Windows (PowerShell):

winget install Docker.DockerDesktop
irm https://astral.sh/uv/install.ps1 | iex
winget install Microsoft.AzureCLI

macOS / Linux:

# Docker: https://docs.docker.com/get-docker/
curl -LsSf https://astral.sh/uv/install.sh | sh
# Azure CLI: https://learn.microsoft.com/cli/azure/install-azure-cli

Verify:

docker --version
uv --version
az account show

• 01_single_agent: Host a single conversational agent and interact with it via a client. Demonstrates basic worker-client architecture and agent state management.
• 02_multi_agent: Host multiple domain-specific agents (physicist and chemist) and route requests to the appropriate agent based on the question topic.
• 03_single_agent_streaming: Enable reliable, resumable streaming using Redis Streams with agent response callbacks. Demonstrates non-blocking agent execution and cursor-based resumption for disconnected clients.

• 04_single_agent_orchestration_chaining: Chain multiple invocations of the same agent using durable orchestration, preserving conversation context across sequential runs.
• 05_multi_agent_orchestration_concurrency: Run multiple agents concurrently within an orchestration, aggregating their responses in parallel.
• 06_multi_agent_orchestration_conditionals: Implement conditional branching in orchestrations with spam detection and email assistant agents. Demonstrates structured outputs with Pydantic models and activity functions for side effects.
• 07_single_agent_orchestration_hitl: Human-in-the-loop pattern with external event handling, timeouts, and iterative refinement based on human feedback. Shows long-running workflows with external interactions.

• 08_workflow: Host a MAF Workflow as a durable orchestration on a standalone worker via DurableAIAgentWorker.configure_workflow. Demonstrates conditional routing and mixing AI agents with non-agent executors.
• 09_workflow_hitl: A workflow that pauses for human approval using ctx.request_info / @response_handler, with the client discovering and answering the pending request.
• 10_workflow_streaming: Stream a hosted workflow's events as typed WorkflowEvent objects by polling the orchestration's custom status.
• 11_subworkflow: Compose workflows by embedding an inner Workflow as a node via WorkflowExecutor. On the durable host the inner workflow runs as its own child orchestration, and a single configure_workflow call registers both.
• 12_subworkflow_hitl: A human-in-the-loop pause that lives inside a sub-workflow. The nested request surfaces to the client with a qualified request id ({executor}~{ordinal}~{requestId}) behind a single top-level addressing surface.

These samples host workflows and agents on Azure Durable Functions (func start) instead of the worker-client model above. Each has its own setup steps in its README, and shared environment setup lives in azure_functions/README.md.

• azure_functions/01_single_agent: Host a single AI agent on Azure Functions with direct HTTP API access for interactive conversations.
• azure_functions/02_multi_agent: Host multiple AI agents on Azure Functions, each reachable via its own HTTP endpoint.
• azure_functions/03_reliable_streaming: Reliable, resumable streaming for durable agents using Redis Streams with cursor-based reconnection.
• azure_functions/04_single_agent_orchestration_chaining: Chain two invocations of the same agent inside a Durable Functions orchestration, preserving conversation state between runs.
• azure_functions/05_multi_agent_orchestration_concurrency: Run two agents in parallel inside a Durable Functions orchestration and merge their responses.
• azure_functions/06_multi_agent_orchestration_conditionals: Conditional orchestration that screens emails with a spam-detector agent and drafts replies with an email assistant agent.
• azure_functions/07_single_agent_orchestration_hitl: Human-in-the-loop orchestration where a writer agent iterates until a reviewer approves or the attempt limit is reached.
• azure_functions/08_mcp_server: Expose agents as both HTTP endpoints and Model Context Protocol (MCP) tools.
• azure_functions/09_workflow_shared_state: Run a MAF Workflow with SharedState on Azure Durable Functions.
• azure_functions/10_workflow_no_shared_state: Run a MAF Workflow on Azure Durable Functions without SharedState.
• azure_functions/11_workflow_parallel: Parallel execution of executors and agents in an Azure Durable Functions workflow.
• azure_functions/12_workflow_hitl: The workflow human-in-the-loop pattern on Azure Durable Functions, with the reviewer notified from inside the workflow via WorkflowHitlContext.
• azure_functions/13_subworkflow_hitl: A human-in-the-loop pause inside a sub-workflow on Azure Durable Functions, exposed through a single top-level respond surface.

These samples are designed to be run locally in a cloned repository.

The following prerequisites are required to run the samples:

• Python 3.9 or later
• Azure CLI installed and authenticated (az login)
• Microsoft Foundry project with a deployed model, configured through FOUNDRY_PROJECT_ENDPOINT and FOUNDRY_MODEL (gpt-4o-mini or better is recommended)
• Durable Task Scheduler (local emulator or Azure-hosted)
• Docker installed if running the Durable Task Scheduler emulator locally

These samples are configured to use the Azure OpenAI service with RBAC permissions to access the model. You'll need to configure the RBAC permissions for the Azure OpenAI service to allow the Python app to access the model.

Below is an example of how to configure the RBAC permissions for the Azure OpenAI service to allow the current user to access the model.

Bash (Linux/macOS/WSL):

az role assignment create \
  --assignee "yourname@contoso.com" \
  --role "Cognitive Services OpenAI User" \
  --scope /subscriptions/<your-subscription-id>/resourceGroups/<your-resource-group-name>/providers/Microsoft.CognitiveServices/accounts/<your-openai-resource-name>

PowerShell:

az role assignment create `
  --assignee "yourname@contoso.com" `
  --role "Cognitive Services OpenAI User" `
  --scope /subscriptions/<your-subscription-id>/resourceGroups/<your-resource-group-name>/providers/Microsoft.CognitiveServices/accounts/<your-openai-resource-name>

More information on how to configure RBAC permissions for Azure OpenAI can be found in the Azure OpenAI documentation.

Most samples use the Durable Task Scheduler (DTS) to support hosted agents and durable orchestrations. DTS also allows you to view the status of orchestrations and their inputs and outputs from a web UI.

To run the Durable Task Scheduler locally, you can use the following docker command:

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.

Each sample reads configuration from environment variables. You'll need to set the following environment variables:

Bash (Linux/macOS/WSL):

export FOUNDRY_PROJECT_ENDPOINT="https://your-project.services.ai.azure.com/api/projects/your-project"
export FOUNDRY_MODEL="your-deployment-name"

PowerShell:

$env:FOUNDRY_PROJECT_ENDPOINT="https://your-project.services.ai.azure.com/api/projects/your-project"
$env:FOUNDRY_MODEL="your-deployment-name"

Navigate to the sample directory and install dependencies. For example:

cd samples/01_single_agent
pip install -r requirements.txt

If you're using uv for package management:

uv pip install -r requirements.txt

Each sample follows a worker-client architecture. Most samples provide separate worker.py and client.py files, though some include a combined sample.py for convenience.

Running with separate worker and client:

In one terminal, start the worker:

python worker.py

In another terminal, run the client:

python client.py

Running with combined sample:

python sample.py

The sample output is displayed directly in the terminal where you ran the Python script. Agent responses are printed to stdout with log formatting for better readability.

You can also see the state of agents and orchestrations in the Durable Task Scheduler dashboard at http://localhost:8082.