Key takeaways
- Reached 1.0 general availability on April 3, 2026 for both .NET and Python, making it the supported successor to AutoGen (now maintenance-mode) and Semantic Kernel — 11.2k+ GitHub stars and weekly releases as of June 2026
- Merges AutoGen's multi-agent patterns (handoff, group chat, Magentic-One) with Semantic Kernel's enterprise features: graph-based workflows with checkpointing, session state management, middleware, and OpenTelemetry observability
- Multi-provider from day one — Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini, and Ollama — plus MCP and A2A protocol support, all MIT-licensed with no commercial tier
FAQ
What is Microsoft Agent Framework?
Microsoft Agent Framework is an open-source SDK and runtime for building AI agents and multi-agent workflows in .NET and Python, created by the same teams as AutoGen and Semantic Kernel as the direct successor to both.
How much does Microsoft Agent Framework cost?
The framework is free and MIT-licensed with no commercial tier. You pay only for the underlying model and cloud services you connect it to, such as Azure OpenAI or Microsoft Foundry.
What models does Microsoft Agent Framework support?
Stable connectors ship for Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini, and Ollama, plus MCP support for tool integration.
How is Microsoft Agent Framework different from AutoGen?
Agent Framework is AutoGen's official successor — it keeps AutoGen's multi-agent orchestration patterns but adds Semantic Kernel's enterprise features (state management, middleware, telemetry) and a graph-based workflow engine; AutoGen itself is now maintenance-mode.
How is Microsoft Agent Framework different from LangChain?
LangChain is Python/TypeScript with the largest integration ecosystem; Agent Framework is .NET + Python with first-class Azure and Microsoft Foundry integration and identical concepts across both languages.
Executive Summary
Microsoft Agent Framework is Microsoft's unified, open-source SDK and runtime for building AI agents and multi-agent workflows in .NET and Python, created by the same teams behind AutoGen and Semantic Kernel as the direct successor to both.[1] Version 1.0 reached general availability on April 3, 2026, marking the framework as production-ready with stable APIs across both languages.[2] The repository stands at 11.2k+ GitHub stars under an MIT license, with the dotnet-1.10.0 release shipping June 10, 2026.[3]
The framework's pitch is consolidation: AutoGen's simple abstractions for single- and multi-agent patterns (handoff, group chat, Magentic-One) combined with Semantic Kernel's enterprise-grade features — session-based state management, type safety, middleware, and telemetry — plus a new graph-based workflow engine for explicit multi-agent orchestration.[1] AutoGen (50K+ stars) is now maintenance-mode and community-managed, receiving only bug fixes and security patches, with Microsoft directing new development to Agent Framework via official migration guides.[4][5]
| Attribute | Value |
|---|---|
| Company | Microsoft |
| Founded | 2025 (announced Oct 2025; 1.0 GA April 3, 2026)[6][2] |
| Funding | Microsoft-backed |
| GitHub Stars | 11.2k+ (as of June 2026)[3] |
| License | MIT[3] |
| Headquarters | Redmond, WA |
Product Overview
Agent Framework splits its surface into two primary capability categories: Agents — individual LLM-powered agents that call tools and MCP servers — and Workflows — graph-based orchestrations connecting agents and functions for multi-step tasks with type-safe routing, checkpointing, and human-in-the-loop support.[1] The docs are explicit about when to use which: agents for open-ended or conversational tasks, workflows for processes with well-defined steps — and "if you can write a function to handle the task, do that instead of using an AI agent."[1]
Key Capabilities
| Capability | Description |
|---|---|
| Agents | LLM-powered agents with tools, MCP servers, and multi-turn sessions |
| Workflows | Graph-based engine with type-safe routing, checkpointing, streaming, human-in-the-loop |
| Orchestration patterns | Sequential, concurrent, handoff, group chat, and Magentic-One (GA)[2] |
| Middleware | Hooks for intercepting and extending agent behavior[2] |
| Memory | Pluggable architecture — Foundry Agent Service, Mem0, Redis, Neo4j, or custom[2] |
| Declarative agents | YAML-defined agents and workflows for configuration management[2] |
| Protocols | MCP and A2A (Agent-to-Agent) support, both GA[2] |
Product Surfaces / Editions
| Surface | Description | Availability |
|---|---|---|
Microsoft.Agents.AI (.NET) | Core framework for .NET | GA (1.0, April 2026)[2] |
agent-framework (Python) | Core framework for Python | GA (1.0, April 2026)[2] |
| DevUI | Browser-based agent debugger | Preview[2] |
| Foundry hosted agents | Managed agent hosting in Microsoft Foundry | Preview[2] |
| Agent Harness | Local agent runtime | Preview[2] |
| Skills | Reusable domain capability packages | Preview[2] |
Technical Architecture
Agent Framework ships the same concepts and APIs across .NET and Python — a deliberate contrast with the Python-only posture of most competing frameworks. The repository itself is 50.7% Python, 46.0% C#, and 2.8% TypeScript (the latter for tooling like DevUI, not an official SDK).[3]
pip install agent-framework
dotnet add package Microsoft.Agents.AI.Foundry --prerelease
Key Technical Details
| Aspect | Detail |
|---|---|
| Deployment | Self-hosted library/runtime; optional Foundry hosted agents (preview) |
| Model(s) | Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini, Ollama[2] |
| Integrations | MCP clients and servers, A2A protocol, AG-UI/CopilotKit/ChatKit adapters (preview), GitHub Copilot SDK and Claude Code SDK integration (preview)[2] |
| Observability | Built-in OpenTelemetry integration for distributed tracing[3] |
| Open Source | Yes (MIT License)[3] |
Foundational Building Blocks
- Model clients — chat completions and responses interfaces[1]
- Agent sessions — state management for multi-turn and long-running scenarios[1]
- Context providers — agent memory abstraction[1]
- Middleware — intercept agent actions for guardrails, logging, policy[1]
Strengths
- Official successor with migration paths — Created by the AutoGen and Semantic Kernel teams; official migration guides exist for both predecessor frameworks, de-risking the transition[1][5]
- True dual-language parity — Same concepts and APIs in .NET and Python at 1.0 GA, rare among agent frameworks[2]
- Proven orchestration patterns at GA — Sequential, concurrent, handoff, group chat, and Magentic-One shipped stable, not experimental[2]
- Multi-provider, not Azure-only — Stable connectors for OpenAI, Anthropic, Bedrock, Gemini, and Ollama alongside Azure OpenAI and Foundry[2]
- Enterprise plumbing built-in — Middleware, session state, checkpointing, pluggable memory (Mem0, Redis, Neo4j), and OpenTelemetry tracing without bolt-ons[2][3]
- Open standards posture — MCP and A2A protocol support at GA, YAML declarative agents for config-driven deployment[2]
- Rapid release cadence — dotnet-1.10.0 shipped June 10, 2026, roughly two months after GA[3]
Cautions
- Young community — 11.2k+ stars as of June 2026 versus 50K+ for AutoGen and far larger numbers for LangChain; the ecosystem of examples, courses, and third-party content is still forming[3][4]
- Migration tax for incumbents — Existing AutoGen and Semantic Kernel codebases must migrate; AutoGen receives only bug fixes and security patches going forward[5][4]
- Preview-heavy edges — DevUI debugger, Foundry hosted agents, Skills, Agent Harness, and the Copilot/Claude Code SDK integrations are all still preview[2]
- No TypeScript/JavaScript SDK — Official SDKs are .NET and Python only, leaving Node.js teams out
- Gravitational pull toward Foundry — The framework is multi-provider, but hosted agents, tools, and the deepest managed features route through Microsoft Foundry[2]
- Third-party risk disclaimer — Microsoft's docs explicitly place responsibility for third-party servers, agents, and non-Azure models on the developer[1]
Pricing & Licensing
| Tier | Price | Includes |
|---|---|---|
| Open Source | Free | Full framework, both .NET and Python SDKs (MIT License)[3] |
Licensing model: Open source (MIT); no commercial tier for the framework itself.[3]
Hidden costs: Model and cloud consumption are separate — Azure OpenAI, Microsoft Foundry, or third-party provider bills scale with usage, and Foundry hosted agents (preview) will carry Azure platform costs.
Competitive Positioning
Direct Competitors
| Competitor | Differentiation |
|---|---|
| LangChain/LangGraph | LangChain has the largest ecosystem (Python + TypeScript, 1000+ integrations); Agent Framework offers .NET parity and native Azure/Foundry depth |
| CrewAI | CrewAI specializes in role-based agent teams with a commercial platform; Agent Framework is a pure framework with Microsoft's enterprise patterns |
| AutoGen | Its own predecessor — AutoGen is maintenance-mode; Agent Framework carries its patterns forward with production hardening[4] |
| OpenAI Agents SDK | OpenAI-centric; Agent Framework is multi-provider across seven stable connectors |
When to Choose Microsoft Agent Framework Over Alternatives
- Choose Agent Framework when: You're a .NET shop, an Azure/Foundry customer, or an existing AutoGen/Semantic Kernel user with a supported migration path
- Choose LangChain when: You want maximum integrations, TypeScript support, and the largest community
- Choose CrewAI when: You want role-based multi-agent teams with a managed enterprise platform
- Choose AutoGen when: Effectively never for new projects — it is maintenance-mode and Microsoft directs new work here[4]
Ideal Customer Profile
Best fit:
- .NET/enterprise development teams — the only major agent framework with first-class C# at GA
- Azure and Microsoft Foundry customers wanting native integration
- Existing AutoGen or Semantic Kernel users facing the migration decision
- Organizations requiring built-in observability, middleware, and state management for compliance
- Teams standardizing on MCP and A2A open protocols
Poor fit:
- Node.js/TypeScript teams (no official SDK)
- Teams wanting a mature third-party ecosystem of tutorials and integrations today
- Organizations avoiding Microsoft ecosystem gravity
- Projects needing managed hosted agents in GA form (Foundry hosting is still preview)
Viability Assessment
| Factor | Assessment |
|---|---|
| Financial Health | Backed by Microsoft — no funding risk |
| Market Position | Challenger — 11.2k+ stars vs. LangChain/CrewAI incumbency, but the default for the Microsoft ecosystem[3] |
| Innovation Pace | Rapid — 1.0 GA April 2026, v1.10.0 by June 10, 2026, large preview-feature pipeline[2][3] |
| Community/Ecosystem | Early — inherits AutoGen/Semantic Kernel communities via migration, still building its own |
| Long-term Outlook | Strong — it is Microsoft's singular, consolidated agent bet |
Microsoft consolidated two overlapping projects into one supported framework and shipped GA in both languages simultaneously — a clear signal of long-term commitment. The open questions are community velocity (can it convert AutoGen's 50K-star community rather than lose it to LangChain or CrewAI) and how much of the preview pipeline hardens to GA over the next year.[4]
Bottom Line
Microsoft Agent Framework is the safest agent framework choice in the Microsoft ecosystem and the only major option with .NET as a first-class citizen. The 1.0 GA delivered real substance — five orchestration patterns, seven model connectors, MCP/A2A protocols, and enterprise plumbing — rather than a rebrand.[2]
Recommended for: .NET and Azure-centric teams building production agents, and any existing AutoGen or Semantic Kernel user — the migration is officially supported and the predecessors are no longer where investment goes.[5]
Not recommended for: TypeScript/Node.js teams, or organizations that want the deepest third-party ecosystem and maximum distance from Microsoft platform gravity.
Outlook: With AutoGen in maintenance mode and Semantic Kernel users being directed to migrate, Agent Framework will absorb most Microsoft-ecosystem agent development by default.[4][5] The framework went from announcement (October 2025) to GA (April 3, 2026) to v1.10.0 (June 2026) in eight months;[6][2][3] the metric to watch is whether its star count and community content close the gap with LangChain and CrewAI now that the consolidation story is settled.
Research by Ry Walker Research • methodology
Sources
- [1] Microsoft Agent Framework Overview (Microsoft Learn)
- [2] Microsoft Agent Framework Version 1.0 (Dev Blogs)
- [3] Microsoft Agent Framework GitHub Repository
- [4] AutoGen GitHub Repository
- [5] Migrate Semantic Kernel and AutoGen Projects to Agent Framework
- [6] Introducing Microsoft Agent Framework (Foundry Blog)