Azure Functions now runs AI agents with markdown config files

Key Takeaways

• Azure Functions now supports AI agents defined in .agent.md markdown files instead of scattered code
• The runtime adds no extra cold start or billing beyond standard function execution
• Agents can connect to 1,400+ enterprise connectors including Teams, Outlook, Salesforce, and ServiceNow

Microsoft turned Azure Functions into an AI agent platform at Build 2026. The serverless agents runtime, now in public preview, lets developers define agents in markdown files rather than code. Any existing Azure Functions trigger, from HTTP to Service Bus to the new Teams and Outlook triggers, can spawn an agent that reasons, calls tools, and connects to enterprise systems.

The billing model stays unchanged. There's no "agents tax." You pay for function execution time with scale-to-zero, identical to running any other workload on Flex Consumption. Cold start? The Azure Functions team told InfoQ the agents runtime adds nothing beyond what you'd see with a regular HTTP trigger. The bottleneck is the LLM, not the infrastructure.

What is the .agent.md format?

The most distinctive choice here is how agents are defined. Where LangChain or CrewAI use Python classes and AutoGen uses TypeScript, Microsoft chose structured markdown. A single .agent.md file contains the agent's system prompt, trigger configuration, available tools, and MCP server connections.

YAML frontmatter declares metadata and the trigger type. The markdown body becomes the agent's instructions. A timer-triggered agent that summarizes tech news and emails it daily looks like this:

---
name: Daily Tech News Email
description: Fetches top tech news and emails a summary daily.
trigger:
  type: timer_trigger
  args:
    schedule: "0 0 15 * * *"
---

You are a news assistant. When triggered, do the following:
1. Scour the web for today's top tech news headlines.
2. Summarize the top stories in a concise, well-formatted HTML email body.
3. Email the summary to $TO_EMAIL with the subject "Daily Tech News Summary" followed by today's date.

That's the entire agent. No Python project with dependencies and boilerplate. MCP servers and tools get declared in companion mcp.json and agents.config.yaml files. The runtime handles orchestration, tool dispatch, and response generation.

Why did Microsoft pick markdown over code?

Readability matters when agents interact with enterprise systems. A CTO reviewing an agent that has access to Salesforce and ServiceNow connectors can read the .agent.md file and understand exactly what it does. No Python knowledge required. The instructions are plain English, and the configuration is standard YAML.

This also makes version control cleaner. The entire agent definition lives in one file. Diffs are readable. Code review is straightforward. Compare that to an agent spread across multiple Python files with decorator-based tool definitions.

What can these agents connect to?

The runtime plugs into Microsoft's connector catalog: 1,400+ managed connectors for Microsoft 365, Teams, Outlook, SharePoint, Salesforce, ServiceNow, and other enterprise systems. Agents also get access to MCP tool servers, sandboxed code execution, and browser automation via Azure Container Apps dynamic sessions.

Built-in surfaces include a chat UI, HTTP chat API, and MCP server endpoint. These are opt-in with no extra code. Authentication runs through managed identity. Traces go to Application Insights. If you've deployed Functions before, the operational model is identical.

How is Microsoft using this internally?

The Azure Functions team built a security audit agent for their own GitHub organizations. Thiago Almeida, Principal Program Manager, described it to InfoQ: the agent runs on a timer trigger, audits branch protection, secret scanning, and workflow permissions across every repository, then reports findings through the same connectors and MCP tool servers available to external developers.

“Because it runs on Flex Consumption, it costs nothing between runs.”

— Thiago Almeida, Principal Program Manager, Azure Functions

This is the serverless value proposition applied to agents. Most agents are idle most of the time. Scale-to-zero billing means you only pay when the agent actually runs.

What else shipped at Build 2026?

The MCP extension reached general availability, expanding from a single tool trigger to full MCP server support with tool, resource, and prompt triggers. It works across .NET, Java, Python, TypeScript, and JavaScript. Built-in On-Behalf-Of authentication lets MCP servers hosted on Functions act on behalf of users without custom auth code.

New connection-backed triggers respond to Teams messages, Outlook mail, calendar events, and SharePoint items. An agent can now wake up when someone sends a Teams message, reason about the content, and take action through any of the 1,400+ connectors.

How does this compare to AWS and Google Cloud?

AWS Lambda and Google Cloud Functions don't have equivalent agent-native runtimes. You can run agent frameworks on them, but you're bringing your own orchestration, tool dispatch, and connector management. Microsoft is betting that first-party agent support, combined with the Microsoft 365 connector ecosystem, creates enough lock-in to matter.

The markdown-first approach is also a bet. Other frameworks assume developers want full programmatic control. Microsoft is gambling that many agent use cases are simple enough that declarative configuration is preferable.

Frequently Asked Questions

Does the Azure Functions agents runtime cost extra?

No. Microsoft confirmed there is no agents tax. Billing is identical to standard function execution on Flex Consumption with scale-to-zero.

What languages does the agents runtime support?

The MCP extension supports .NET, Java, Python, TypeScript, and JavaScript. The .agent.md format itself is language-agnostic.

Can Azure Functions agents connect to non-Microsoft services?

Yes. The 1,400+ connector catalog includes Salesforce, ServiceNow, and other third-party services alongside Microsoft 365, Teams, and SharePoint.

Is the serverless agents runtime generally available?

No. It launched in public preview at Build 2026. The MCP extension reached GA, but the agents runtime is still preview.

Source: InfoQ