Microsoft Agent Framework for orchestrated agents

How it works

The framework provides agent abstractions and a workflow layer for coordinating steps and carrying state. The hosting application decides where the agent runs and which dependencies it receives.

What you need

Start with a language-specific sample and a configured model provider. Define one well-scoped tool that the workflow can call without sharing application-wide credentials.

What to try first

Start with one agent and one deterministic tool. Add a workflow step that reviews the result and test the behavior when that step fails or needs input.

Use fixed responses in a workflow unit test. This makes it easier to debug orchestration and retries without attributing a failure to model variability.

Where to look for failures

Language SDKs and provider integrations can differ. Verify the runtime and service requirements for the sample you choose, and define ownership of retries across workflow and tool boundaries.

Related Franklin pages

Build the same two-step workflow with one tool and one failure case. Inspect how state, errors and tool permissions cross the boundary between the steps.

• Expose application functions to models through plugins and provider connectors. Semantic Kernel for integrating models into applications
• Create agents and workflows with tools, session state and development utilities. Google ADK for Python agent applications

Browse Franklin for concrete examples of tools and services exposed through MCP. Franklin’s MCP directory

Project documentation

• Current project README
• README at the reviewed revision