Azure Functions Introduces Markdown-First Serverless Agents Runtime

At Build 2026, Microsoft expanded Azure Functions into a serverless agents platform, enabling developers to define AI agents in markdown rather than imperative code. The shift reflects a broader bet on declarative, infrastructure-agnostic agent authoring.

The .agent.md Model#

Agents are now described in single .agent.md files combining YAML frontmatter (trigger config, metadata) and markdown instructions. A timer-triggered news summarizer, for instance, needs just one file—no Python boilerplate or scattered configuration:

---
name: Daily Tech News Email
trigger:
  type: timer_trigger
  args:
    schedule: "0 0 15 * * *"
---
You are a news assistant...

Tool declarations live in companion mcp.json and agents.config.yaml files. The runtime parses and orchestrates everything.

Agents Spawn from Any Trigger#

HTTP, Timer, Service Bus, Event Hubs, SQL, Cosmos DB, and new Microsoft 365 triggers (Teams messages, Outlook mail, SharePoint items) can now instantiate agents. Each trigger type has first-class support without extra code.

No Cold Start or Cost Tax#

Microsoft confirmed that agents don't add latency overhead beyond standard HTTP triggers on Flex Consumption—the bottleneck is the LLM, not infrastructure. Billing is identical to regular functions: per-second, scale-to-zero, no premium for agents.

Broader Runtime Upgrades#

• MCP Extension (GA): Full Model Context Protocol support across .NET, Java, Python, TypeScript, JavaScript with built-in On-Behalf-Of auth inheritance
• Durable Task Scheduler: Now validates scale; runs hundreds of millions of executions weekly for complex AI workflows
• On-Demand Sandboxes (private preview): Isolated microVMs per orchestration step for native toolchains (ffmpeg, LibreOffice), CPU-heavy tasks, and customer plugin safety
• Go as first-class language, Flex Consumption rolling updates (GA), container-based OS dependencies