Microsoft shipped its Agent Framework as an open-source AI agent SDK on June 19, 2026. The release appeared on Microsoft and GitHub under an MIT license. The SDK lets developers build, test, and deploy agentic workflows without paying per-agent fees. It ships with Python and .NET packages, a local runtime, and optional Azure AI Foundry hooks. The package is around 2.4 MB and carries no model weights, so there is no parameter count. That small footprint matters for teams watching the AI API free tier limits that tightened across major providers in June 2026. Microsoft framed this as a production-grade release, not a research preview.
Why it matters: this is Microsoft’s first production-grade open-source agent SDK that separates framework code from model billing. Unlike closed agent features in Copilot, the Agent Framework can run fully local with Ollama, llama.cpp, or any OpenAI-compatible endpoint. Microsoft confirmed the MIT license on its vendor announcement page, not a hidden repo. The SDK also includes four built-in modules: memory, tool orchestration, guardrails, and evaluation. That means a team can move from a prototype to a governed deployment without stitching together LangChain, LangSmith, and cloud safety filters. It lands as major AI model tier changes push free users toward lighter models, making a model-agnostic runtime more useful.
Microsoft released the Agent Framework on June 19, 2026, under the MIT license. It is not a model and has no context window of its own. The framework passes through model context limits, including 1 million token windows on compatible models like Gemini 2.5 Pro and GPT-5.5. Benchmarks are provided through a built-in evaluation harness that includes GAIA and AgentBench-style task templates with pass@1 scoring. The vendor did not claim a single model benchmark because the SDK does not ship weights. That is a key distinction for open-source readers who track free AI models with no API costs.
The timing is sharp. Microsoft Copilot locked some Office apps behind a paywall in June 2026, and agentic AI billing changes hit free users hard. The Agent Framework arrives as a lower-cost alternative for developers who want to own the orchestration layer and swap models later. Because the SDK is open source, teams can audit the guardrails, extend memory, and avoid vendor-specific agent surcharges. Microsoft said the initial release supports Python 3.11 and 3.12 plus .NET 10, with a TypeScript preview on the roadmap.
How Do the Top Options Compare?
| Framework | Best For | License | Language Support | Built-in Guardrails |
|---|---|---|---|---|
| Microsoft Agent Framework | Model-agnostic orchestration | MIT | Python, .NET | Yes: jailbreak, prompt injection, content safety |
| OpenAI Agents SDK | OpenAI-first workflows | Apache 2.0 | Python, TypeScript | Yes: OpenAI moderation |
| LangGraph | Graph-based agent pipelines | MIT | Python, JavaScript | Via extensions |
| CrewAI | Role-based multi-agent teams | MIT | Python | Prompt-level only |
| AutoGen | Conversational multi-agent research | MIT | Python, .NET | Research-level only |
License details and feature availability are based on vendor documentation as of June 19, 2026. Guardrail support varies by model provider and deployment target.
1. Microsoft Agent Framework , Best for model-agnostic agent orchestration
Microsoft Agent Framework 1.0 shipped on June 19, 2026 under the MIT license. The SDK is model-agnostic and supports OpenAI, Anthropic, Google, Meta, DeepSeek, and local runtimes. It includes four modules: memory, tools, guardrails, and evaluation. The Python package is about 2.4 MB and installs via pip. There is no parameter count because the framework ships no weights. The package is available on GitHub and model cards for supported open-weight models appear on Hugging Face. It also works with Microsoft Azure AI Foundry for managed deployments.
The memory module stores conversation state, short-term context, and long-term vector search. Developers can plug in FAISS, Pinecone, or Azure AI Search. The tool orchestration layer supports function calling, MCP servers, and code execution. Guardrails include content safety, jailbreak detection, and prompt injection filters. Evaluation templates include GAIA, AgentBench, and SWE-bench task formats with pass@1 and step-count metrics.
The framework matters because it removes per-agent runtime fees. You do not pay Microsoft when you run agents locally. You only pay if you call a hosted model or use Azure services. That changes the math for teams burned by recent agentic AI billing changes. The SDK is not a Copilot feature and does not require Microsoft 365. It is designed to be audited, self-hosted, and swapped between model providers.
Key strengths:
- ✅ Open-source MIT license allows commercial use and modification
- ✅ Model-agnostic runtime supports local, OpenAI, Anthropic, Google, and DeepSeek models
- ✅ Built-in guardrails cover prompt injection, jailbreak, and content safety
- ✅ Evaluation harness includes GAIA and AgentBench-style tasks
- ✅ Python and .NET packages ship at production quality
- ❌ No official TypeScript support in the 1.0 release
- ❌ Managed deployment beyond Azure requires self-hosting skills
- ❌ Memory and tool modules still rely on external providers for vector search
Who it’s for: Choose this if you want an open-source agent runtime with guardrails and do not want to pay per-agent orchestration fees.
2. OpenAI Agents SDK , Best for OpenAI-first agent workflows
OpenAI’s Agents SDK is the company’s official framework for building agents around GPT-5.5, GPT-5.5-mini, and o-series models. It ships with Python and TypeScript support and uses the same API abstraction as the Responses API. The license is Apache 2.0, but the framework is optimized for OpenAI endpoints. You can connect other providers through compatible endpoints, but the developer experience favors OpenAI.
The SDK includes handoffs, guardrails, sessions, and tracing. OpenAI uses built-in content moderation and safety classifiers. Pricing is separate: you pay for model tokens and any agent tracing storage. That can create confusion when free tier limits shift, as seen in AI API free tier limits.
OpenAI Agents SDK is easier for teams already using ChatGPT or the Responses API. It lacks the local-first posture of Microsoft Agent Framework. If you want to run models entirely offline, this is not the strongest choice. The framework also ties you more closely to OpenAI’s model release cadence, which changed several times in June 2026.
Key strengths:
- ✅ Native support for GPT-5.5 and o-series reasoning models
- ✅ Trace dashboard is simple to use
- ✅ TypeScript and Python packages are well documented
- ❌ OpenAI-first design can lock you into one vendor
- ❌ Local model support is limited
- ❌ Agent tracing can add hidden costs
Who it’s for: Choose this if you already use OpenAI models and want the fastest path to production agents.
3. LangGraph , Best for complex graph-based agent pipelines
LangGraph is an open-source orchestration framework from LangChain. It uses a graph abstraction where nodes are steps and edges define control flow. The MIT-licensed library supports Python and JavaScript. It is popular for stateful, human-in-the-loop workflows that need branching, retries, and time travel debugging.
LangGraph has a large ecosystem of integrations and a paid platform, LangSmith, for evaluation and monitoring. The framework itself is free, but production features like tracing often push teams toward paid tiers. Microsoft Agent Framework ships evaluation built in, which reduces the need for a separate paid observability tool. That distinction matters for teams navigating free AI pricing changes.
LangGraph offers more graph flexibility than Microsoft’s SDK, but that flexibility comes with a steeper learning curve. Teams focused on production guardrails may prefer a framework that includes eval and safety filters without a separate subscription. LangGraph remains a strong choice for complex stateful agents, especially when you already use LangChain.
Key strengths:
- ✅ Graph-based control flow handles complex multi-step agents
- ✅ Large integration library
- ✅ Python and JavaScript support
- ❌ Advanced monitoring and eval often require LangSmith paid plans
- ❌ Steeper learning curve than linear agent SDKs
- ❌ Guardrails are not built into core
Who it’s for: Choose this if you need graph based workflows and already use LangChain components.
4. CrewAI , Best for role-based multi-agent teams
CrewAI is an open-source framework for role playing multi-agent crews. You define agents with roles, goals, and backstories, then assign tasks. The MIT-licensed Python library runs locally and supports many model providers. CrewAI also has a paid platform for deployment and monitoring.
CrewAI is simpler than LangGraph and good for demos or internal automations. Its role based model can create impressive outputs but can be harder to govern in production. Microsoft Agent Framework includes content safety and jailbreak guardrails by default, which is a stronger fit for enterprise use. CrewAI’s guardrails are mostly prompt-level and require manual work.
CrewAI has a growing community and many templates. It lacks the built-in evaluation harness that Microsoft’s SDK offers. Teams dealing with Anthropic agent billing changes may want to keep orchestration costs separate from model costs, which both frameworks allow.
Key strengths:
- ✅ Role based design is easy to understand
- ✅ Large template library
- ✅ MIT license
- ❌ Production guardrails are limited
- ❌ Evaluation tools require separate setup
- ❌ Python only for core development
Who it’s for: Choose this if you want simple role based multi-agent automation without deep graph logic.
5. AutoGen , Best for conversational multi-agent research
AutoGen is Microsoft’s earlier open-source framework for conversational multi-agent systems. It is MIT-licensed and supports Python and .NET. AutoGen is widely used in research and prototyping, but its production hardening has lagged. Microsoft positions the new Agent Framework as the more governed successor, with built-in safety and evaluation.
AutoGen lets multiple agents chat to solve problems, which is flexible and great for experimentation. It lacks the turnkey guardrails and evaluation harness of the Agent Framework. Teams can migrate by keeping their AutoGen tool definitions and moving orchestration to the new SDK. This reduces the cost of switching without discarding existing agent logic.
AutoGen remains a strong choice for exploratory work and academic projects. For production deployments that touch user data, the Agent Framework’s guardrails matter more. The shift mirrors free AI pricing changes that push teams toward self-hosted, cost-controlled stacks.
Key strengths:
- ✅ Flexible conversational agent patterns
- ✅ MIT license
- ✅ Python and .NET support
- ❌ Production safety features are minimal
- ❌ Evaluation tools are not built in
- ❌ Overlapping with the newer Agent Framework can confuse adopters
Who it’s for: Choose this if you are researching conversational multi-agent systems and need maximum flexibility.
Frequently Asked Questions
What is Microsoft Agent Framework?
Microsoft Agent Framework is an open source SDK for building, testing, and deploying AI agents. It ships under the MIT license with Python and .NET support. The 1.0 release includes memory, tool orchestration, guardrails, and evaluation modules. It does not include model weights and has no parameter count.
Is Microsoft Agent Framework free to use?
Yes. The framework is free under the MIT license. You can run it locally without paying Microsoft. You only incur costs if you call a hosted model provider or use managed Azure services.
Which models work with Microsoft Agent Framework?
The framework is model agnostic. It supports OpenAI, Anthropic, Google, Meta, DeepSeek, and any OpenAI-compatible endpoint. It also runs local models through Ollama or llama.cpp. Context window support depends on the chosen model.
Does Microsoft Agent Framework require Azure?
No. Azure is optional. The SDK runs locally on Linux, macOS, and Windows. You can add Azure AI Foundry for managed deployments, but no Azure account is required for local use.
How does Microsoft Agent Framework compare to LangGraph?
Microsoft Agent Framework includes guardrails and evaluation built in, while LangGraph relies on extensions or LangSmith for some of that. LangGraph offers more flexible graph based control flow. Microsoft’s SDK is more prescriptive and production oriented out of the box.
What license does Microsoft Agent Framework use?
Microsoft Agent Framework ships under the MIT license. This allows commercial use, modification, and distribution. The license applies to the SDK, not to any underlying models you connect.
What Should You Remember?
- Microsoft Agent Framework shipped on June 19, 2026 under an MIT license with Python and .NET support.
- No model weights means the SDK has no parameter count and requires an external or local model.
- Model-agnostic lets developers swap between OpenAI, Anthropic, Google, Meta, DeepSeek, and local runtimes.
- Built-in guardrails cover jailbreak, prompt injection, and content safety without extra vendor fees.
- Evaluation harness includes GAIA and AgentBench-style tasks with pass@1 scoring.
- No Azure required for local development, but Azure AI Foundry is optional for managed deployment.
- Lower cost compared to closed agent features in Copilot or per-agent orchestration billing.
Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.
