microsoft-agent-framework

Microsoft Agent Framework guidance for building, updating, refactoring, explaining, and reviewing agent-based solutions in either .NET or Python. Positioned as the unified successor to Semantic Kernel and AutoGen, the framework is in public preview and evolves rapidly, so the central discipline is grounding every recommendation in current official documentation and samples rather than potentially stale knowledge. The first decision point is language selection: .NET workflows apply when a repository contains .cs, .csproj, .sln, or .slnx files (or the user asks for C#), while Python workflows apply for .py, pyproject.toml, or requirements.txt. Mixed repositories follow the files being edited, and ambiguous cases prompt an inspection of the workspace before choosing a language-specific reference. Once the language is set, contributors consult the matching reference file for package names, repository paths, and sample locations. Shared cross-language practices include async patterns for agent and workflow operations, explicit error handling and logging, strong typing with clear interfaces, and DefaultAzureCredential for Azure authentication. The skill distinguishes two building blocks: agents handle autonomous decision-making, ad hoc planning, conversation flows, tool usage, and MCP server interactions, while workflows handle multi-step orchestration, predefined execution graphs, long-running tasks, and human-in-the-loop scenarios. It supports multiple model providers such as Azure AI Foundry, Azure OpenAI, and OpenAI, while favoring Azure AI Foundry for new projects, and applies advanced patterns like thread-based state, context providers, middleware, checkpointing, routing, and orchestration where they fit. Migration is a first-class concern: existing Semantic Kernel or AutoGen code is treated as input to be ported using the official migration guides, preserving behavior first and adopting native patterns incrementally. When repository examples diverge from current documentation, the difference is explained and the supported pattern is followed. Use this when starting new agent projects, modernizing legacy orchestration code, or reviewing existing Agent Framework applications for alignment with current best practices.