Microsoft Agent Framework 1.0 ships graph workflows and MCP, with A2A next

Microsoft shipped Microsoft Agent Framework 1.0 for .NET and Python, positioning it as a production-ready release with stable APIs and long-term support. It is also a consolidation moment: Microsoft is explicitly framing this SDK as the successor path for teams coming from AutoGen and Semantic Kernel.

Primary sources: Microsoft’s 1.0 announcement (devblogs.microsoft.com/agent-framework/microsoft-agent-framework-version-1-0/) and the Microsoft Learn overview (learn.microsoft.com/agent-framework/overview).

What shipped in 1.0 (stable surface)

Microsoft lists the 1.0 feature surface as battle-tested and committed to backward compatibility, including stable single agents/connectors, middleware hooks, memory/context providers, graph workflows, and multi-agent orchestration patterns.

• Workflows: graph-based composition of agents and functions, with checkpointing and hydration.
• Orchestration patterns: sequential, concurrent, handoff, group chat, plus Magentic-One.
• Interop: MCP tool discovery and invocation; A2A called out with “coming soon” language.

Why “graph-based workflows” is the real engineering upgrade

The interesting part is not “multi-agent” as a vibe. It is that the workflow topology is explicit: you can branch on conditions, fan out work in parallel, and converge results in a repeatable way. Microsoft also calls out checkpointing and hydration so long-running runs survive interruptions.

Practitioner payoff: explicit graphs are easier to test, easier to resume after failures or approvals, and easier to observe and govern with policy middleware.

Interop: MCP now, A2A soon

Microsoft calls out cross-runtime interoperability via MCP and A2A. MCP is described as the mechanism for discovering and invoking external tools exposed by MCP-compliant servers. A2A is referenced, but the 1.0 post also says A2A 1.0 support is “coming soon”, so treat it as roadmap until the integration is versioned and documented.

AutoGen + Semantic Kernel: successor framing is explicit

Microsoft Learn calls Agent Framework the direct successor and the next generation of both Semantic Kernel and AutoGen. The announcement positions 1.0 as the production-ready point after an RC period where the feature surface was locked and validated with community feedback.

What is preview (functional, still evolving)

Microsoft lists several preview features since the October introduction. Treat these like preview APIs: useful, but likely to change.

• DevUI debugger for visualizing execution and tool calls
• Foundry hosted agent integration (managed services, Azure Durable Functions)
• AG-UI adapters and frontend streaming integrations (CopilotKit, ChatKit)
• Skills packages; Copilot SDK and Claude Code SDK integrations; “agent harness” runtime

Decision rules for teams

Microsoft Learn’s framing is simple and practical: use an agent for open-ended conversational tasks, and a workflow for well-defined multi-step processes where you need explicit control over execution order and coordination.

• Adopt MCP early if tool portability is a priority and you can expose tools via MCP servers.
• Be cautious with A2A until “coming soon” becomes a versioned, documented integration with compatibility guarantees.

Getting started

Microsoft’s announcement highlights: pip install agent-framework (Python) and dotnet add package Microsoft.Agents.AI (.NET). Start with the announcement, the Learn overview, and the GitHub repo: github.com/microsoft/agent-framework.