congiuluc/agent-framework-python

Microsoft Agent Framework — Python

When to Use

• Creating or modifying AI agents using Microsoft Agent Framework in Python
• Adding function tools, MCP tools, or code-interpreter capabilities to agents
• Building multi-turn conversational agents with session state
• Implementing graph-based workflows with executors and edges
• Orchestrating multi-agent patterns (sequential, concurrent, hand-off)
• Adding middleware for logging, security, error handling on agent runs
• Integrating with Azure OpenAI, OpenAI, Anthropic, Ollama, or other providers
• Streaming agent responses
• Migrating from Semantic Kernel or AutoGen to Agent Framework

Official Documentation

• Agent Framework Overview
• Your First Agent
• Add Tools
• Multi-Turn Conversations
• Agents Overview
• Function Tools
• Tool Approval (Human-in-the-loop)
• Agent Middleware
• Sessions & State
• Context Providers
• Workflows Overview
• Providers
• Integrations (A2A, AG-UI, Azure Functions, M365)
• Migration from Semantic Kernel
• Migration from AutoGen
• GitHub Repository — Python
• PyPI Package

Installation

pip install agent-framework --pre

This installs all sub-packages. For individual packages, see python/packages in the GitHub repo.

Procedure

1. Identify the capability needed (single agent, tools, multi-turn, workflow, middleware)
2. Follow the patterns in Python agent guidelines
3. Apply the code style rules
4. Review the matching sample in ./samples/
5. Use @kernel_function decorator or plain functions to define tools
6. Use agent sessions for multi-turn conversations — never manage raw chat history
7. Use middleware for cross-cutting concerns (logging, security, error handling)
8. For multi-agent patterns, use graph-based Workflows with typed executors and edges
9. Use async def for all agent invocations with proper error handling
10. Use AzureCliCredential during development; prefer ManagedIdentityCredential in production
11. Add OpenTelemetry instrumentation for observability
12. Add Google-style docstrings to all public classes and functions
13. Create corresponding tests (see testing-backend-python skill)