microsoft/agent-framework

Microsoft Agent Framework is an open-source toolkit from Microsoft for building AI agents and systems where multiple AI agents work together. An AI agent, in this context, is a program that can take instructions in plain language, reason through a task, use tools, and produce a result, often through multiple steps without constant human direction. The framework is designed to take these agents from early experiments into production systems that actually run reliably. The framework supports both Python and C# (.NET), so teams can use whichever language their project already uses. It works with several AI providers, including Azure OpenAI, OpenAI, and Microsoft's own Foundry infrastructure, and is designed to let you switch between them without rewriting your entire application. One of the central features is orchestration: the ability to connect multiple agents together into workflows. These workflows can be sequential (one agent passes its result to the next), concurrent (multiple agents work in parallel), or more complex patterns where agents hand off tasks to each other or collaborate in a group. The framework includes checkpointing, which means a workflow can be paused and restarted without losing progress, and human-in-the-loop support, which lets a person review or approve a step before it continues. For observability, the framework integrates with OpenTelemetry, a standard tool for tracking what a distributed system is doing at runtime, which matters when debugging or monitoring agents in production. Agents can also be defined using YAML files rather than code, which can make configuration easier to version and share. Installation is a single pip command for Python or a dotnet add package command for .NET. Microsoft provides documentation, tutorials, and migration guides for teams moving from other agent frameworks like Semantic Kernel or AutoGen.