# Microsoft Agent Framework

> Source: https://aiwiki.ai/wiki/microsoft_agent_framework
> Updated: 2026-07-28
> Fact-checked: 2026-07-28
> Categories: AI Agents, Developer Tools, Microsoft
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Microsoft Agent Framework." aiwiki.ai, 28 Jul 2026. https://aiwiki.ai/wiki/microsoft_agent_framework
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Microsoft Agent Framework** (MAF) is an open-source software development kit and runtime developed by [Microsoft](https://aiwiki.ai/wiki/microsoft) for building [AI agents](https://aiwiki.ai/wiki/ai_agents), multi-agent systems, and graph-based workflows. It provides related programming models for .NET and Python, while a separate Go implementation is in public preview. The framework is not a language model, a hosted agent service, or an end-user assistant: applications built with it still depend on a chosen model provider, tools, data stores, and deployment environment.[1][2]

Microsoft announced MAF in public preview on October 1, 2025, as a convergence of ideas and components from [AutoGen](https://aiwiki.ai/wiki/autogen) and [Semantic Kernel](https://aiwiki.ai/wiki/semantic_kernel). Both .NET and Python reached release-candidate status on February 19, 2026. Microsoft recorded general availability on April 2 and published the version 1.0 announcement on April 3.[1][3][4][5] Version 1.0 established stable core APIs and a long-term support commitment for the two main implementations, but individual integrations and extension packages have their own maturity levels.[4]

The project is released under the [MIT license](https://aiwiki.ai/wiki/mit_license). Microsoft distributes the Python implementation through packages headed by `agent-framework` and the .NET implementation through the `Microsoft.Agents.AI` package family. The Go module is developed in a separate Microsoft repository and does not yet have full feature parity with the .NET and Python SDKs.[6][24]

## History

MAF followed two Microsoft projects with different origins. AutoGen began as a [Microsoft Research](https://aiwiki.ai/wiki/microsoft_research) project on programmable conversations among agents; its initial paper appeared in 2023.[16] Semantic Kernel was a model-agnostic application SDK with state, plugins, connectors, and enterprise-oriented integration features. Microsoft described MAF as combining AutoGen's agent and orchestration concepts with Semantic Kernel's sessions, type system, middleware, telemetry, and provider support.[2]

| Date | Milestone |
| --- | --- |
| October 1, 2025 | Microsoft introduced MAF in public preview as an open-source SDK and runtime for multi-agent systems.[1] |
| February 19, 2026 | The .NET and Python implementations entered release-candidate status. Microsoft said the intended 1.0 feature set was complete and the API surface was stable.[3] |
| April 2-3, 2026 | MAF reached version 1.0 for .NET and Python. Microsoft recorded GA on April 2 and published its announcement on April 3.[4][5] |
| June 3, 2026 | Microsoft presented additional preview and integration work at Build 2026, including the agent harness, Foundry hosted agents, and CodeAct.[5] |
| July 8, 2026 | The Python orchestration package reached 1.0, placing the standard orchestration patterns at 1.0 in both main SDKs.[21] |
| July 22, 2026 | Microsoft released the core agent harness as stable for .NET and Python.[22] |
| July 23, 2026 | Declarative workflow support reached 1.0 across .NET and Python.[23] |

These dates do not imply that every feature became stable with the base 1.0 release. For example, the April announcement listed DevUI, hosted-agent integration, skills, and the agent harness among features that were still in preview at that time. Several of those components moved to stable releases later in July, while some optional harness tools remained unreleased as of July 22.[4][22]

## Architecture

Microsoft's current documentation divides MAF into three capability groups: agents, the agent harness, and workflows. Shared building blocks include model clients, sessions, context providers, middleware, tool integrations, and event or telemetry hooks.[2]

### Agents and state

An agent wraps a model client with instructions, tools, and runtime behavior. The core interface is named `AIAgent` in .NET, `BaseAgent` in Python provider APIs, and `agent.Agent` in Go. Provider adapters expose a broadly consistent interface, but their capabilities differ. The documented provider matrix, for example, shows that file search, code interpretation, MCP tools, structured outputs, and background responses are not available through every adapter.[10]

`AgentSession` carries conversation state between runs. A session can hold local state and, for some providers, an identifier for history stored by a remote service. Context and history providers can add past messages, retrieved information, user preferences, or application-specific state. Sessions can be serialized and restored, but persistence depends on the provider or storage integration selected by the application.[8]

Middleware can inspect or transform model calls, agent runs, and tool invocations without placing all cross-cutting logic in a prompt. Typical uses include logging, policy checks, retries, telemetry, and approval handling. This makes the framework extensible, but does not make those controls automatic: an application has to configure the relevant middleware and enforce its own policy.

### Workflows and orchestration

A workflow is a directed graph whose executors perform work and whose edges route typed messages. The runtime emits events while the graph runs and supports streaming, shared state, checkpointing, and human-in-the-loop pauses. Its execution model groups runnable executors into synchronized supersteps. Executors within a superstep can run concurrently, while a barrier separates one superstep from the next and provides a checkpoint boundary.[7]

The lower-level graph API can connect agents with ordinary functions. Higher-level [agent orchestration](https://aiwiki.ai/wiki/agent_orchestration) builders produce common coordination patterns: sequential, concurrent, group chat, handoff, and magentic orchestration. The magentic pattern is derived from work including [Magentic-One](https://aiwiki.ai/wiki/magentic_one); a manager agent plans and revises work among specialist agents. These builders return ordinary workflows, so an application can combine a standard pattern with custom executors and routing.[21]

Declarative workflows describe orchestration in YAML rather than application control flow. They can express state changes, conditions, loops, jumps, function or agent calls, MCP and HTTP actions, approval points, and checkpoint or resume behavior. The YAML is loaded into the same workflow type used by code-first definitions. This can make orchestration easier to review and version, but the referenced agents, credentials, tools, and policies still have to be supplied by the host application.[23]

### Agent harness

The harness is an opinionated configuration for long, multi-step tasks. Its stable core includes an automatic tool-calling loop, per-service-call history persistence, context compaction, planning and todo state, file-backed memory, skills, tool approvals, provider-backed web search when available, and OpenTelemetry instrumentation.[22]

The harness does not make an agent autonomous in an unlimited sense. It executes within configured iteration limits, tool permissions, storage boundaries, and approval rules. At the July 22 release, background subagents, direct file access, automatic looping, and shell tooling were still opt-in features that Microsoft had not marked as released. Their availability should not be inferred from the stable status of the core harness.[22]

## Tools and interoperability

MAF supports three broad tool paths. Function tools call application code; MCP clients connect to tools exposed by [Model Context Protocol](https://aiwiki.ai/wiki/model_context_protocol) servers; provider-hosted tools use capabilities such as web search, file search, or code interpretation offered by an inference service. Tool definitions are included in model context, so large tool catalogs add token overhead and can reduce selection reliability. A tool call is still a model-generated request that the application or service executes, not an action performed by the language model itself.[9]

MCP and Agent2Agent (A2A) serve different boundaries. MCP defines client-server messages for capabilities such as tools, resources, and prompts, using JSON-RPC as its base message protocol. A2A defines discovery and task exchange among independent agent systems, including synchronous, streaming, and asynchronous interactions.[11][12] MAF can consume MCP tools and expose or consume agents through A2A, but protocol support does not guarantee that two providers expose identical model, tool, authentication, or session features.

For hosting, Microsoft documents A2A endpoints, OpenAI-compatible chat or response endpoints, AG-UI connections for web interfaces, and a Durable Extension for long-running work. The Durable Extension can persist sessions and workflow checkpoints on Azure Functions or self-hosted workers. It is distinct from Microsoft Foundry Agent Service, which is a separate managed hosting service.[13][14]

MAF emits observability data through OpenTelemetry. Traces can describe model operations, tool calls, and workflow activity, and can be exported to compatible backends. OpenTelemetry's generative AI semantic conventions remain a separate standard; using its vocabulary does not by itself provide evaluation, access control, or data-retention policy.[15]

## Relationship to earlier projects

Microsoft calls MAF the direct successor to both AutoGen and Semantic Kernel, but the earlier projects do not have the same maintenance status.

| Project | Origin and role | Status as of July 28, 2026 |
| --- | --- | --- |
| AutoGen | Microsoft Research framework for applications built from conversations among configurable agents.[16] | In maintenance mode and community managed. Its repository says it will receive bug fixes, security patches, and documentation work rather than new features, and directs new users to MAF.[17] |
| Semantic Kernel | Microsoft SDK for model integration, plugins, memory, and agent applications across .NET, Python, and Java. | Still supported as the 1.x line. Microsoft committed to critical fixes, security work, and selected feature completion while moving most new agent features to MAF. Its repository continued to publish releases in July 2026.[18][19] |
| Microsoft Agent Framework | Successor framework that combines concepts from both projects with a graph workflow runtime. | Stable at the core 1.x level for .NET and Python. The Go implementation remains in public preview and has documented feature gaps.[2][4][24] |

Migration therefore does not mean that AutoGen or Semantic Kernel code is automatically compatible with MAF. Microsoft maintains separate migration guides, and applications must map project-specific agent types, state, tools, and orchestration behavior to the new APIs. In particular, describing Semantic Kernel as retired is inaccurate: Microsoft's stated policy was to support it for the foreseeable future and for at least one year after MAF reached GA.[18]

AG2 is a separate community project that evolved from AutoGen and announced its own organization and governance in November 2024. Its maintainers describe it as volunteer maintained and use Apache License 2.0 for later modifications, while preserving the original licensing of inherited code. AG2 is not another name for Microsoft's current AutoGen repository or for MAF.[20]

## Licensing and support

The main MAF repository and the Go implementation use the MIT license.[6][24] The framework can connect to Microsoft services and to third-party providers, including [Amazon Bedrock](https://aiwiki.ai/wiki/amazon_bedrock), Anthropic, Google, Ollama, and OpenAI, but the exact connector set and feature matrix vary by language and release.[4][10]

Microsoft's repository support policy draws a boundary between the open-source SDK and Azure service support. Customers with a Unified Support agreement can receive Microsoft support when an issue arises from using the SDK with Azure AI services. Uses that do not involve an Azure service are treated as use of an open-source tool and are directed to GitHub or community forums.[26] This is narrower than saying that every MAF deployment receives enterprise product support.

## Security and limitations

MAF is a development framework, not a safety boundary or a completed AI system. Its transparency documentation says that accuracy, helpfulness, and safety depend on the selected models and the application built around them. Model outputs can still be inaccurate or biased, and agents can choose the wrong tool, supply invalid arguments, or expose data to an external service.[25] General generative AI risks identified by NIST, including confabulation, information-security failures, harmful content, and human over-reliance, continue to apply.[27]

The framework provides mechanisms that can support safer operation, including scoped tools, human approval, middleware, checkpointing, telemetry, and isolated code execution through selected integrations. These mechanisms have to be configured for the use case. Microsoft's tool guidance recommends narrowly defined operations rather than unrestricted database or shell access, approval before consequential actions, error handling, and independent verification where needed.[9]

Session identifiers also require application-level authorization. Microsoft warns that provider-issued conversation or response identifiers should not be treated as end-user authorization boundaries when one service credential is shared among users. A multi-tenant host should keep provider identifiers in trusted storage, map them to its own client-facing identifiers, and verify the authenticated user or tenant before resuming a session.[8]

Connections to non-Microsoft models, MCP servers, agents, or code can move data outside an organization's expected compliance or geographic boundary. The provider documentation places responsibility on the application operator to review those data flows, permissions, retention rules, and third-party terms.[10] Observability and workflow durability can improve diagnosis and recovery, but neither establishes that an agent's decisions are correct.

## See also

- [LangChain](https://aiwiki.ai/wiki/langchain)
- [LangGraph](https://aiwiki.ai/wiki/langgraph)
- [Agent orchestration](https://aiwiki.ai/wiki/agent_orchestration)
- [Model Context Protocol](https://aiwiki.ai/wiki/model_context_protocol)
- [Magentic-One](https://aiwiki.ai/wiki/magentic_one)

## References

1. Yina Arenas, "Introducing Microsoft Agent Framework," Microsoft Azure Blog, October 1, 2025. https://azure.microsoft.com/en-us/blog/introducing-microsoft-agent-framework/
2. Microsoft, "Microsoft Agent Framework Overview," Microsoft Learn, updated July 10, 2026. https://learn.microsoft.com/en-us/agent-framework/overview/
3. Shawn Henry, "Microsoft Agent Framework Reaches Release Candidate," Microsoft Foundry Blog, February 19, 2026. https://devblogs.microsoft.com/foundry/microsoft-agent-framework-reaches-release-candidate/
4. Shawn Henry, "Microsoft Agent Framework Version 1.0," Microsoft Agent Framework Blog, April 3, 2026. https://devblogs.microsoft.com/agent-framework/microsoft-agent-framework-version-1-0/
5. Shawn Henry, "Microsoft Agent Framework at Build 2026: Agent Harness, Hosted Agents, CodeAct, and more," Microsoft Agent Framework Blog, June 3, 2026. https://devblogs.microsoft.com/agent-framework/microsoft-agent-framework-at-build-2026-announce/
6. Microsoft, "microsoft/agent-framework," GitHub repository and license, accessed July 28, 2026. https://github.com/microsoft/agent-framework
7. Microsoft, "Workflow Builder and Execution," Microsoft Learn, updated July 10, 2026. https://learn.microsoft.com/en-us/agent-framework/workflows/workflows
8. Microsoft, "Session," Microsoft Learn, accessed July 28, 2026. https://learn.microsoft.com/en-us/agent-framework/agents/conversations/session
9. Microsoft, "Adding Tools," Microsoft Learn, accessed July 28, 2026. https://learn.microsoft.com/en-us/agent-framework/journey/adding-tools
10. Microsoft, "Providers Overview," Microsoft Learn, accessed July 28, 2026. https://learn.microsoft.com/en-us/agent-framework/agents/providers/
11. Model Context Protocol contributors, "Model Context Protocol Specification, revision 2025-11-25," accessed July 28, 2026. https://modelcontextprotocol.io/specification/2025-11-25/basic
12. A2A Protocol Working Group, "Agent2Agent Protocol Specification 1.0.0," accessed July 28, 2026. https://a2a-protocol.org/latest/specification/
13. Microsoft, "Host Your Agent," Microsoft Learn, accessed July 28, 2026. https://learn.microsoft.com/en-us/agent-framework/get-started/hosting
14. Microsoft, "Durable Extension," Microsoft Learn, accessed July 28, 2026. https://learn.microsoft.com/en-us/agent-framework/integrations/durable-extension
15. OpenTelemetry, "Generative AI semantic conventions," accessed July 28, 2026. https://opentelemetry.io/docs/specs/semconv/gen-ai/
16. Qingyun Wu et al., "AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation," arXiv:2308.08155, 2023. https://arxiv.org/abs/2308.08155
17. Microsoft, "microsoft/autogen," GitHub repository, accessed July 28, 2026. https://github.com/microsoft/autogen
18. Shawn Henry, "Semantic Kernel and Microsoft Agent Framework," Microsoft Agent Framework Blog, October 7, 2025. https://devblogs.microsoft.com/agent-framework/semantic-kernel-and-microsoft-agent-framework/
19. Microsoft, "microsoft/semantic-kernel releases," GitHub, accessed July 28, 2026. https://github.com/microsoft/semantic-kernel/releases
20. AG2 contributors, "ag2ai/ag2," GitHub repository, accessed July 28, 2026. https://github.com/ag2ai/ag2
21. Evan Mattson, "Agent Framework's Orchestration Patterns Reach 1.0," Microsoft Agent Framework Blog, July 8, 2026. https://devblogs.microsoft.com/agent-framework/agent-frameworks-orchestration-patterns-reach-1-0/
22. Wes Steyn, "The Microsoft Agent Framework Harness is now released," Microsoft Agent Framework Blog, July 22, 2026. https://devblogs.microsoft.com/agent-framework/the-microsoft-agent-framework-harness-is-now-released/
23. Peter Ibekwe, "Move Agent Orchestration/Workflows out of Code with Agent Framework Declarative Workflows 1.0," Microsoft Agent Framework Blog, July 23, 2026. https://devblogs.microsoft.com/agent-framework/move-agent-orchestration-workflows-out-of-code-with-agent-framework-declarative-workflows-1-0/
24. Microsoft, "microsoft/agent-framework-go," GitHub repository, accessed July 28, 2026. https://github.com/microsoft/agent-framework-go
25. Microsoft, "Microsoft Agent Framework Responsible AI Transparency FAQs," GitHub, accessed July 28, 2026. https://github.com/microsoft/agent-framework/blob/main/TRANSPARENCY_FAQ.md
26. Microsoft, "Microsoft Agent Framework Support Policy," GitHub, accessed July 28, 2026. https://github.com/microsoft/agent-framework/blob/main/SUPPORT.md
27. National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile," NIST AI 600-1, July 2024. https://doi.org/10.6028/NIST.AI.600-1