Microsoft Agent Framework

Microsoft Agent Framework is Microsoft's open-source (MIT-licensed) SDK for building AI agents and multi-agent workflows, and the direct successor to both Semantic Kernel and AutoGen. Microsoft's documentation states that it "combines AutoGen's simple agent abstractions with Semantic Kernel's enterprise features — session-based state management, type safety, middleware, telemetry — and adds graph-based workflows for explicit multi-agent orchestration", calling it "the next generation of both Semantic Kernel and AutoGen", created by the same teams. Version 1.0 for .NET and Python was announced on April 3, 2026 as a "production-ready release: stable APIs, and a commitment to long-term support."

Status and Languages

As of October 6, 2026 the framework is generally available for .NET and Python, with a Go implementation in public preview; Microsoft's overview notes that declarative agents, RAG, CodeAct and functional workflows are not yet available in Go. The two GA languages release independently and frequently: the newest tags on GitHub are dotnet-1.23.0 (October 1, 2026) and python-1.20.0 (October 2, 2026). Migration guides exist for both predecessors.

The framework is not tied to Azure-hosted models. The 1.0 announcement lists first-party connectors for Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini and Ollama. It is, however, the most complete agent framework for the .NET ecosystem, which is its clearest point of difference from Python-first alternatives such as LangGraph, CrewAI and the OpenAI Agents SDK.

Agents and Workflows

The framework separates two ways of building. An agent is a model with instructions, tools and a session, suited to open-ended or conversational work where the model decides the steps. A workflow connects agents and ordinary functions through explicit execution paths, suited to processes with well-defined steps where the developer wants control over ordering. Microsoft's own guidance on choosing between them is unusually direct: "If you can write a function to handle the task, do that instead of using an AI agent."

Area	What it provides
Agents	Single agents that call tools and MCP servers across multiple model providers
Workflows	Functional and graph-based workflows of executors and edges, with human-in-the-loop pauses and checkpoints
Orchestrations	Prebuilt sequential, concurrent, handoff, group-chat and Magentic multi-agent patterns
Building blocks	Agent sessions for state, context providers for memory, middleware for intercepting agent actions, telemetry
Harness Agent	An opinionated agent for long tasks with planning, todo tracking, context compaction, file access and tool approval

The prebuilt orchestrations cover the common multi-agent shapes, including the orchestrator–worker pattern, and a workflow can itself be exposed through the standard agent interface. Agents and workflows can also be defined declaratively in YAML and kept under version control. The 1.0 announcement listed several components as still in preview at that date, among them the DevUI debugger, Foundry hosted integration, AG-UI adapters, Skills packages, the GitHub Copilot and Claude Code SDK integrations, and the Agent Harness.

Checkpointing

Durability is the feature that most distinguishes the workflow engine from a simple agent loop. Workflows execute in supersteps, and a checkpoint is written at the end of each one, after every executor in that superstep has finished. According to the documentation a checkpoint captures the state of all executors, the messages pending for the next superstep, pending requests and responses, and shared state. A run can be restored to any earlier checkpoint, or a new workflow instance can be rehydrated from one, provided it has the same topology and executor identities as the workflow that wrote it.

Microsoft lists the intended uses as recovering long-running workflows after failure, pausing and resuming, periodic state saving for audit or compliance, and migrating a run between environments. In Python the built-in stores are in-memory, local file and Azure Cosmos DB; from Python 1.13.0 the engine also writes an entry checkpoint before the first superstep and when responses to request events are delivered, which makes a complete run replayable. Combined with human-in-the-loop pauses, this lets a workflow stop for an approval, persist, and continue later, a building block of agent reliability.

The documentation is explicit about one risk: checkpoint storage "is a trust boundary". The Python file and Cosmos stores serialize non-JSON state with pickle, mitigated by a restricted unpickler that allows only a set of safe types by default, and Microsoft warns never to load checkpoints from untrusted or potentially tampered sources.

MCP and A2A

Both open agent protocols are supported in the stable release. Agents can call tools exposed by Model Context Protocol servers, and the framework includes MCP clients among its foundational building blocks.

A2A support runs in both directions. An agent can discover and invoke a remote A2A agent as if it were local, and an Agent Framework agent can be exposed as an A2A server that publishes an agent card at the well-known discovery path and accepts messages over the protocol's HTTP+JSON or JSON-RPC bindings. Hosting is documented for .NET (ASP.NET Core packages under Microsoft.Agents.AI.Hosting.A2A), Python (agent-framework-a2a) and Go. One caveat: as of the documentation dated July 2026, the install commands for the A2A hosting packages still carry prerelease flags, so GA of the core framework does not imply that every integration package is stable.

Assessment

The framework ends the overlap between Semantic Kernel and AutoGen in Microsoft's agent tooling. Its strengths are first-class .NET support, checkpointed graph workflows and standard protocols. The trade-offs are a large surface area, a fast release cadence across two independently versioned languages, and uneven maturity between the stable core and the preview packages around it. Independent, published evidence comparing its production behaviour with other frameworks is thin; most available material is Microsoft's own. A structured comparison is at Microsoft Agent Framework vs LangGraph.