Microsoft Agent Framework: Open-Source Engine for Agentic AI App (devblogs.microsoft.com)

Microsoft announced the open-source Microsoft Agent Framework, a unified SDK and runtime that merges the enterprise stability of Semantic Kernel with the experimental multi-agent orchestration of AutoGen. It’s designed to take agent prototypes into production by providing a single foundation for building, deploying, and managing multi-agent systems—supporting both LLM-driven “agent orchestration” for creative reasoning and deterministic “workflow orchestration” for repeatable business processes. The framework emphasizes portability and standards (MCP, A2A, OpenAPI), cloud-agnostic runtime options, and tighter dev tooling (VS Code AI Toolkit updates, Microsoft 365 Agents SDK integration) so teams can prototype locally and scale to Azure AI Foundry without rewrites. Technically, Agent Framework is modular and extensible: declarative agent definitions (YAML/JSON), pluggable memory backends (Redis, Pinecone, Qdrant, Weaviate, Elasticsearch, Postgres, etc.), built-in connectors (Microsoft Graph, Fabric, SharePoint, Oracle, Amazon Bedrock, MongoDB), and support for OpenAPI-imported tools. It combines deterministic and dynamic orchestration patterns (sequential, concurrent, group chat, handoff, “magentic” manager-agent coordination) and exposes enterprise features—OpenTelemetry observability, long-running durability with pause/resume, content safety, Entra ID auth, human-in-the-loop approvals, and CI/CD integrations. For AI/ML teams this lowers the friction from research to production, standardizes cross-agent interoperability, and provides the governance and reliability enterprises require for real-world agent deployments.