MicrosoftMicrosoft Agent Framework (MAF) is a multi-language framework for building, orchestrating, and operating AI agents and multi-agent workflows. It ships consistent APIs for Python and C#/.NET, with a Go SDK maintained in a separate repository. It is the successor to Semantic Kernel and AutoGen, and migration guides cover moving from either.
Agents connect to a range of LLM providers, including Microsoft Foundry, Azure OpenAI, and OpenAI, and the framework also works with the GitHub Copilot SDK. Multi-agent systems are built as graph-based workflows with sequential, concurrent, handoff, and group collaboration patterns, plus checkpointing, streaming, human-in-the-loop control, and time travel.
The framework targets agents meant to run in production, where durability, restartability, observability, and governance matter. Agents can be defined in code or declaratively in YAML, hosted locally, exposed over A2A (agent-to-agent), or deployed to Foundry-hosted infrastructure. A separate Durable Agent Framework extension adds Durable Task and Azure Functions hosting. A developer UI (DevUI) supports interactive testing and debugging, and experimental AF Labs packages cover benchmarking and reinforcement learning.
Features
- Python and .NET: consistent APIs across Python and C#/.NET, with a separate Go SDK
- Multiple providers: Microsoft Foundry, Azure OpenAI, OpenAI, and other LLM providers
- Workflows: graph-based multi-agent orchestration with sequential, concurrent, handoff, and group collaboration patterns
- Durable runs: checkpointing, streaming, human-in-the-loop control, and time travel inside workflows
- Middleware: request and response processing, exception handling, and custom processing chains
- Foundry hosted agents: deploys agents to Foundry-hosted infrastructure with two extra lines of code
- Observability: built-in OpenTelemetry tracing, monitoring, and debugging
- Declarative agents: defines agents in YAML for setup and versioning
- Agent skills: domain knowledge from files, inline code, and class libraries that agents discover and use
- DevUI: an interactive developer UI for building, testing, and debugging agents and workflows
- AF Labs: experimental packages for benchmarking, reinforcement learning, and research
