Serverless AI agents with Azure Functions

This sample shows how to build and deploy AI agents on Azure Functions using the Azure Developer CLI (azd). It creates a Python function app with two agents:

• main.agent.md: a chat agent with built-in chat UI/API/MCP endpoints and Python dynamic session pool code execution.
• daily_microsoft_blog_summary.agent.md: a timer-triggered agent that summarizes recent Microsoft blog posts with the same dynamic session pool.

Email delivery is optional. If you provide an email recipient, the deployment creates an Office 365 Outlook Connector Gateway and MCP server so the timer agent can email the digest. If you leave the recipient blank, no Office 365 resources are created and the timer agent returns the digest in its final response so you can verify the run in Function logs or Application Insights.

Prerequisites

• An Azure account with an active subscription
• Azure Developer CLI
• Azure Functions Core Tools
• Python 3.13
• Azurite, if you want to run the function app locally

Initialize the project

Clone this repository or initialize it from the template repository:

azd init --template Azure-Samples/functions-quickstart-serverless-agents-azd

Supply an environment name when prompted. In azd, the environment keeps deployment state and is used in the generated resource group name.

Configure optional email delivery

Email delivery is disabled by default. To deploy the sample without Office 365 Outlook, do not set TO_EMAIL.

To enable email delivery, set TO_EMAIL before running azd up:

azd env set TO_EMAIL you@example.com

When email delivery is enabled, azd up provisions an Office 365 Outlook connection and MCP server. After deployment, authenticate the connection in the Connector Namespace portal before expecting the timer agent to send email.

Open the Connector Namespace portal from the deployed environment:

CONNECTOR_PORTAL_URL="https://connectors.azure.com/$(az account show --query id -o tsv)/rg-$(azd env get-value AZURE_ENV_NAME)/$(azd env get-value O365_CONNECTOR_GATEWAY_NAME)/overview"
echo "$CONNECTOR_PORTAL_URL"
open "$CONNECTOR_PORTAL_URL"

Use the connectors.azure.com URL for authorization, not the generic Azure portal resource URL.

Deploy to Azure

Run:

azd up

The deployment creates:

• A Flex Consumption Azure Functions app
• A user-assigned managed identity
• A storage account and Application Insights
• A Microsoft Foundry account, project, and gpt-4.1 deployment
• An Azure Container Apps dynamic session pool
• Optional Office 365 Outlook Connector Gateway resources, when TO_EMAIL is set

The scaffold defaults to gpt-4.1, FOUNDRY_DEPLOYMENT_CAPACITY=200, and no reasoning settings. If you intentionally upgrade to a reasoning-capable model such as gpt-5.4, set the Foundry model parameters and AZURE_FUNCTIONS_AGENTS_REASONING_EFFORT/AZURE_FUNCTIONS_AGENTS_REASONING_SUMMARY together. If your subscription has less remaining quota than the default capacity, lower FOUNDRY_DEPLOYMENT_CAPACITY or choose another region/SKU/model. Use reasoning effort medium by default; raise it to high and increase capacity when the agent needs deeper reasoning.

After deployment, azd stores output values in your local environment. You can review them with:

azd env get-values

Use the chat agent

After azd up completes, open the built-in chat UI:

https://<function-app-name>.azurewebsites.net/agents/main/

The chat agent also exposes POST /agents/main/chat, POST /agents/main/chatstream, and an MCP tool through /runtime/webhooks/mcp. It can use Python code execution through the dynamic session pool, but it does not have access to the Office 365 Outlook email tool.

Verify the timer agent

The timer agent runs once per day by default. To verify timer runs, inspect Function logs or Application Insights for Daily Microsoft Blog Summary Agent responses.

When TO_EMAIL is blank, the timer agent does not try to send email. It returns the complete digest as its final response so the digest appears in logs.

When TO_EMAIL is set and the Office 365 connection is authenticated, the timer agent sends the digest to that recipient using the Office 365 Outlook MCP tool.

Run locally

Copy the sample local settings file and fill in values from a deployed environment:

cd src
cp local.settings.json.sample local.settings.json

For local runs with Microsoft Foundry, set AZURE_FUNCTIONS_AGENTS_PROVIDER to foundry, fill in FOUNDRY_PROJECT_ENDPOINT and FOUNDRY_MODEL, and set ACA_SESSION_POOL_ENDPOINT if you want Python code execution. Set TO_EMAIL, O365_MCP_SERVER_URL, and optionally O365_MCP_CLIENT_ID only when you want local email delivery.

Install dependencies and start the Functions host:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
func start

For local MCP access to Office 365 Outlook, sign in with Azure CLI and make sure your user account has access to the Connector Gateway connection:

az login

Project structure

azure.yaml
infra/
  main.bicep
  main.parameters.json
  app/
    api.bicep
    connector-gateway.bicep
    foundry.bicep
    rbac.bicep
    session-pool.bicep
    session-pool-rbac.bicep
src/
  function_app.py
  main.agent.md
  daily_microsoft_blog_summary.agent.md
  agents.config.yaml
  host.json
  mcp.json
  requirements.txt

Clean up resources

When you are finished, delete the Azure resources:

azd down