Microsoft Agent Framework 1.0

Microsoft Agent Framework (MAF) is the unified SDK that merged AutoGen and Semantic Kernel into a single, production-supported platform. It reached 1.0 GA on April 3, 2026, with the same concepts and APIs across .NET and Python.

For three years, Microsoft ran parallel bets: Semantic Kernel handled model connectors, memory, and tool calling; AutoGen handled multi-agent conversation patterns. Teams had to pick one or stitch them together with custom glue. MAF ends that split. Semantic Kernel becomes the foundation layer; AutoGen-style orchestration sits on top as a graph workflow engine. One install, stable APIs, long-term support.

• You are building agents in the Microsoft ecosystem (.NET, Azure, Microsoft 365).
• You need multi-agent orchestration with explicit graph-based workflows.
• You want a single framework that handles both single-agent tool calling and multi-agent collaboration.
• You are already invested in Semantic Kernel or AutoGen and want the migration path.

• Unified API surface. Chat clients, tools, MCP integrations, context providers, middleware, and multi-step workflows in one programming model. Same concepts across .NET and Python.
• Graph-based workflows. Explicit control over multi-agent execution paths — not just conversation patterns, but structured state-machine workflows with branching, parallelism, and human-in-the-loop.
• Session-based state management. Long-running and human-in-the-loop scenarios are first-class. State persists across steps, not just in-memory.
• MCP support. Native Model Context Protocol integration for connecting agents to external tools and data sources.
• Enterprise features inherited from Semantic Kernel. Type safety, filters, telemetry, middleware, extensive model and embedding support.
• BUILD 2026 additions. Agent Harness, Hosted Agents, and CodeAct were previewed, adding managed execution and code-based agent actions.
• Open source. MIT-licensed, with active development from the same teams that built AutoGen and Semantic Kernel.

• .NET and Azure-native bias. MAF is the obvious default for .NET and Azure teams. For Python-first or polyglot teams already on LangGraph or CrewAI, the calculus is less clear. The Python parity is good but the ecosystem gravity is .NET.
• Migration effort. If you have existing AutoGen or Semantic Kernel code, the migration path exists but is not zero-effort. The workflow API is new and requires rewriting orchestration logic.
• Young ecosystem. Despite the 1.0 label, community content, tutorials, and patterns are still maturing compared to LangGraph's deeper community base. Expect to read source code more often than blog posts.
• Azure-hosted agents add cost. The framework is free, but Hosted Agents and Agent Harness (BUILD 2026 previews) run on Azure infrastructure with associated costs. Budget for compute, not just the SDK.
• Python and .NET feature parity is not always simultaneous. New features may land in one language first. Check the release notes for your target language.

The framework itself is free and open source under MIT. You pay for:

• Compute infrastructure if you self-host (VMs, containers, AKS).
• Azure-hosted agents if you use the BUILD 2026 Hosted Agents preview (usage-based Azure pricing).
• Model API costs from whatever provider you connect (OpenAI, Azure OpenAI, or others via MCP).

Teams already in the Microsoft ecosystem who want a single, supported framework for both single-agent and multi-agent workflows. Especially strong for .NET shops that previously had to choose between Semantic Kernel's enterprise features and AutoGen's multi-agent patterns. For Python-first teams evaluating frameworks, compare directly against LangGraph 1.x and the OpenAI Agents SDK — the right choice depends on your existing stack and workflow complexity.