Microsoft Agent Framework: A Technical Overview
The Microsoft Agent Framework is an open-source SDK and runtime for building and orchestrating production-grade AI agents and multi-agent systems. It consolidates two of Microsoft’s agent projects — AutoGen and Semantic Kernel — into a single pro-code toolchain, available for both Python and .NET, with a managed production path through the Azure AI Foundry Agent Service.
Unifying AutoGen and Semantic Kernel
The framework’s core value is collapsing two previously separate toolchains into one consistent API surface, taking the strengths of each.
| Source | Contribution |
|---|---|
| AutoGen | The agent runtime and multi-agent patterns — abstractions for single- and multi-agent conversations, agent lifecycles, and collaborative behavior. |
| Semantic Kernel | The enterprise layer — thread-based state, type-safe plugins, telemetry and observability hooks, and broad model and embedding support. |
The result lets you build autonomous agents while enforcing the reliability, safety, and manageability that production demands.
Key concepts
Dual orchestration modes
The framework supports two modes natively, used alone or combined in hybrid systems.
| Mode | How the flow is decided | Fits |
|---|---|---|
| Agent orchestration | LLM-driven — agents reason, choose tools, and plan next steps toward a goal. | Open-ended research, dynamic planning, creative problem-solving. |
| Workflow orchestration | Deterministic — business logic hands tasks off in a fixed sequence. | Process automation, data pipelines, rule-based systems. |
Runtime and stateful threads
A managed runtime handles agent lifecycles, identities, and communication, formalizing patterns first prototyped in AutoGen. The thread is the unit of state: it makes conversations reproducible, durable, and auditable — the foundation for debugging, retries, and compliance.
Functions and plugins
Tool use follows Semantic Kernel’s plugin model. External tools — APIs, code interpreters, custom functions — bind to agents through strongly-typed contracts, so capabilities are easy to manage and reuse across agents and workflows.
Model flexibility
The core AIAgent interface is model-agnostic, so you can swap providers without rewriting orchestration logic and tune cost against performance. Supported providers include Azure OpenAI Service, OpenAI’s APIs, GitHub Models, and local runtimes such as Ollama and Foundry Local.
Production on Azure AI Foundry
The SDK is open source, but it targets scalable deployment on the Azure AI Foundry Agent Service, which supplies the runtime and controls production needs:
- Execution — compute for running agents and workflows.
- State — durable thread state for long-running tasks.
- Observability — telemetry, logging, and diagnostics via OpenTelemetry hooks.
- Enterprise controls — identity, networking policy, and content-safety filters.
- Integration — a bridge to the wider Azure AI model catalog and toolchains.
Why it lowers cost
The framework targets the real economics of enterprise AI — token spend, latency, and failure recovery — in three ways. A single unified runtime removes most of the brittle “glue code” that drives development and maintenance cost. Integrated telemetry and stateful threads make latency and failures traceable, cutting operational overhead. And model flexibility lets teams route each task to the most cost-effective model in a multi-agent workflow.
Audience and interoperability
This is a pro-code framework for developers building complex, custom systems — positioned above low-code platforms like Copilot Studio. The AIAgent interface also interoperates with other agent standards, including Azure AI Foundry Agents, OpenAI Assistants, and Copilot Studio, which reduces lock-in and eases integration into a broader stack.
Relationship to AutoGen and Semantic Kernel
The framework is the official successor to both projects, built by the same core teams. Microsoft recommends it for new projects. Existing AutoGen projects keep receiving maintenance — bug fixes and security patches — but new feature work is focused on the unified framework.
