Edit

Compare the serverless agents runtime with other Microsoft agent options

Microsoft offers several ways to build and run AI agents. This article compares the Azure Functions serverless agents runtime with other Microsoft agent options so that you can choose the option that fits a given agent, or combine more than one.

These options differ mainly in the authoring model, how agents start, where they run and bill, and who the primary user is. They also share building blocks: the developer options can call the same Azure OpenAI and Foundry models, several use Microsoft Agent Framework, and the Model Context Protocol (MCP) provides a common way to share tools across them.

Important

The serverless agents runtime is currently in preview. Features, configuration names, and supported connectors can change before general availability.

Compare the options

The following table summarizes the key differences between the ways you can create and host your agents:

Option Authoring model Hosting and cost Best suited for...
Serverless agents runtime Markdown .agent.md files plus Python app Azure Functions hosted with scale-to-zero Event-driven agents that react to triggers
Foundry prompt agents Declarative config in portal or code Managed by Foundry with no agent compute Agents that don't include custom code or require custom hosting
Foundry hosted agents Your code (containerized) Managed by Foundry (container compute) Agents with custom code running in a managed container
Copilot Studio Low-code designer and natural language Microsoft-hosted (Copilot credits) Conversational and autonomous agents for Microsoft 365
Microsoft Scout Natural-language prompts Local (on a user device) Personal task automation on a local device

These options aren't mutually exclusive. For example, a serverless agents runtime app can expose tools through MCP that Foundry agents or Copilot Studio agents consume.