# Microsoft Agent Framework: Production AI Agents in Python, .NET, and Go > Microsoft Agent Framework (MAF) is an open-source framework for building, orchestrating, and deploying AI agents and multi-agent workflows in Python and C#/.NET, with a separate Go SDK. It provides graph-based orchestration patterns, durability, human-in-the-loop control, and OpenTelemetry integration for teams taking agents from prototype to production. **microsoft/agent-framework** — A framework for building, orchestrating and deploying AI agents and multi-agent workflows with support for Python and .NET. - Repository: https://github.com/microsoft/agent-framework - Website: https://aka.ms/agent-framework - Stars: 13,866 · Forks: 2,397 - Language: Python - License: MIT - Published: 2026-09-10 · Updated: 2026-09-10 · Language: en - Canonical page: https://hysenlabs.com/projects/microsoft-agent-framework ## What Microsoft Agent Framework is built for The README's fit assessment is specific: MAF is appropriate when a team is building agents and workflows they expect to run in production, needs orchestration beyond a single prompt or stateless chat loop, wants graph-based patterns including sequential, concurrent, handoff, and group collaboration, cares about durability, restartability, observability, governance, or human-in-the-loop control, and needs provider flexibility so the architecture can evolve without major rewrites. Those five criteria mark the boundary between MAF and simpler agent frameworks. A stateless question-answering loop that calls one LLM endpoint does not need MAF. A multi-agent pipeline that runs document reviews, spawns specialized sub-agents, checkpoints progress, and requires human approval at specific decision points fits the framework's design. MAF supports multiple LLM providers: Microsoft Foundry, Azure OpenAI, OpenAI, and the GitHub Copilot SDK are mentioned in the README, with the note that more providers are being added. This provider flexibility is one of the explicit design goals: the README states the framework keeps architecture choices open as requirements evolve. ## Installing and getting started with the framework The Python package is on PyPI: ```bash pip install agent-framework ``` The README notes that this installs the standard package set. The experimental `agent-framework-lab` package, which covers benchmarking, reinforcement learning, and research features, must be installed separately. On Windows, the first install may take a minute. For .NET, the core package and common Foundry integration packages are: ```bash dotnet add package Microsoft.Agents.AI dotnet add package Microsoft.Agents.AI.Foundry dotnet add package Azure.AI.Projects dotnet add package Azure.Identity ``` The Go SDK is in a separate repository at microsoft/agent-framework-go. The main repository notes that Go documentation, samples, contribution guidance, and the issue tracker are in that repository. The official documentation is at learn.microsoft.com/agent-framework. The README links to an overview, a quick start, tutorials, and a user guide. Migration guides for teams coming from AutoGen or Semantic Kernel are also documented at learn.microsoft.com. ## Orchestration patterns and workflow features MAF's orchestration layer supports four graph patterns documented in the README: sequential (agents run in order), concurrent (agents run in parallel), handoff (one agent delegates to another), and group collaboration (multiple agents work on a shared problem). These patterns map to the common multi-agent architectures in production systems. Beyond basic patterns, the framework supports checkpointing (saving workflow state for restartability), streaming (incremental output delivery), human-in-the-loop control (pausing a workflow for human review), and time-travel (replaying from a previous checkpoint). The README describes these as built-in features, accessible through the workflow API. Foundry Hosted Agents is a feature that deploys agents to Microsoft Foundry-hosted infrastructure. The README states this requires only two additional lines of code beyond a local agent definition. Sample code for Python and .NET Foundry hosting is in the python/samples/04-hosting/ and dotnet/samples/04-hosting/ directories of the repository. Observability is built in through OpenTelemetry. The README links to observability samples in python/samples/02-agents/observability/ and .NET telemetry samples in dotnet/samples/02-agents/AgentOpenTelemetry/, giving teams a path to distributed tracing and monitoring without building custom instrumentation. ## Declarative agents and middleware MAF supports defining agents in YAML for faster setup and versioning. The declarative-agents/ directory in the repository holds examples. YAML-defined agents reduce the amount of Python or C# code needed for agents with straightforward configurations and make the agent definition reviewable as a configuration file. The middleware system provides a pipeline for request and response processing, exception handling, and custom processing steps. The README links to Python middleware samples in python/samples/02-agents/middleware/ and .NET middleware samples in dotnet/samples/02-agents/Agents/Agent_Step11_Middleware/. This is the extension point for cross-cutting concerns like logging, rate limiting, or custom retry logic. Agent Skills allow building domain-specific knowledge bases from multiple sources: files, inline code, and class libraries. According to the README, agents can discover and use these knowledge bases at runtime. The architecture document for skills is at docs/decisions/0037-agent-skills-design.md in the repository. AF Labs is the experimental package namespace for features under development, including benchmarking and reinforcement learning, in the python/packages/lab/ directory. ## The DevUI and the development workflow MAF ships a DevUI, an interactive developer interface for building, testing, and debugging agents and workflows without writing test harness code. The README links to a YouTube video demonstrating the DevUI in action. The DevUI is particularly useful during the orchestration design phase, when the developer needs to see how agents hand off to each other, where human-in-the-loop pauses occur, and how the workflow graph executes. This reduces the iteration time compared to running full integration tests for each change. The repository's latest releases as of the time of writing are python-1.19.0 (published 2026-09-18) and dotnet-1.22.0 (published 2026-09-18). The Python and .NET SDKs are versioned independently, reflecting their separate release cadences. Both are on the public package repositories (PyPI and NuGet) at their current versions. ## Microsoft Agent Framework versus AutoGen The repository provides a migration guide from AutoGen at learn.microsoft.com/en-us/agent-framework/migration-guide/from-autogen, which signals that MAF is intended as a successor to AutoGen within the Microsoft open-source ecosystem. AutoGen is Microsoft's earlier multi-agent framework, focused on conversational multi-agent patterns and automated task completion through agent conversations. The key difference is that MAF is designed around graph-based workflows with explicit patterns (sequential, concurrent, handoff, group), while AutoGen's original design centered on conversational loops between agents. MAF also has more explicit production infrastructure features: checkpointing, observability, and hosted deployment. A migration guide also exists for Semantic Kernel, Microsoft's other AI framework, which is focused on AI plugin composition for single-agent scenarios. Teams using Semantic Kernel for basic plugin orchestration and needing multi-agent coordination may find MAF the appropriate next step. ## License, maintenance, and the repository structure MAF is MIT licensed. The last push to the repository was on 2026-09-27. The repository is actively maintained with recent releases: python-1.19.0 and dotnet-1.22.0 both from 2026-09-18. The repository layout organizes code by language: python/ for the Python SDK with packages and samples, dotnet/ for the C#/.NET SDK, declarative-agents/ for YAML agent examples, and docs/ for architecture decision records. The go/ directory exists in the repository but the README notes that the Go SDK documentation and issue tracker are in the separate microsoft/agent-framework-go repository. The TRANSPARENCY_FAQ.md file in the repository is an unusual artifact: it documents design decisions, comparisons to other frameworks, and rationale for architectural choices. This is worth reading for teams evaluating MAF against other agent frameworks, as it provides first-person reasoning from the project team rather than external analysis. ## Conclusion Microsoft Agent Framework is the right choice for teams building AI agents in Python or C#/.NET that need production-grade features: graph-based orchestration, checkpointing, human-in-the-loop control, and Azure deployment. Teams prototyping single-agent chat loops that do not need durability or multi-agent coordination will find the framework more setup than necessary at that stage. Go users must use the separate microsoft/agent-framework-go repository. Before adopting MAF, check the migration guides for AutoGen and Semantic Kernel: both are documented in the repository and indicate that MAF is positioned as the successor to those frameworks within the Microsoft ecosystem. ## FAQ ### What is Microsoft Agent Framework? Microsoft Agent Framework (MAF) is an open-source Python and C#/.NET framework for building production-grade AI agents and multi-agent workflows. It provides graph-based orchestration patterns, checkpointing, human-in-the-loop control, and Azure deployment support. ### How do I install Microsoft Agent Framework? For Python, run `pip install agent-framework`. For .NET, run `dotnet add package Microsoft.Agents.AI`. The Go SDK is in the separate microsoft/agent-framework-go repository. ### Why use Microsoft Agent Framework? MAF is designed for production AI systems that need graph-based multi-agent orchestration, durability through checkpointing, human-in-the-loop control, and OpenTelemetry observability. It supports Azure OpenAI, OpenAI, and Microsoft Foundry as providers. ### Is Microsoft Agent Framework open source? Yes. The repository is MIT licensed and publicly available at github.com/microsoft/agent-framework. The Go SDK is in a separate repository at github.com/microsoft/agent-framework-go, also open source. ## Sources - [License: MIT](https://github.com/microsoft/agent-framework/blob/main/LICENSE) - [microsoft/agent-framework on GitHub](https://github.com/microsoft/agent-framework) - [Project website](https://aka.ms/agent-framework) - [README](https://github.com/microsoft/agent-framework/blob/main/README.md) - [Releases](https://github.com/microsoft/agent-framework/releases) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/microsoft-agent-framework