Microsoft Agent Framework
Microsoft Agent Framework merges Semantic Kernel and AutoGen into one framework for single and multi-agent apps in Python and .NET, with workflows and enterprise controls.
Overview
Microsoft Agent Framework is an open-source framework for building AI agents and multi-agent workflows in Python and .NET. It is the successor to two earlier Microsoft projects. Semantic Kernel was an enterprise SDK for adding AI to applications. AutoGen was a research framework for agents that talk to each other. Agent Framework combines the stable foundations of the first with the multi-agent ideas of the second, in one codebase.
It offers two main building blocks. An agent is a single model-driven worker that reasons, calls tools and replies. A workflow is a graph that connects agents and ordinary functions in a defined order, with support for branching, parallel steps, checkpoints and human approval. Around these sit middleware, telemetry, connectors to many model providers, and support for the MCP and A2A protocols.
An everyday comparison is a hospital ward. Each nurse and doctor is an agent with a skill set. The ward's care pathway is the workflow: admit, assess, treat, review, discharge, with set points where a senior doctor must sign off. The pathway can be paused overnight and resumed in the morning because everything is written in the patient's chart (a checkpoint). Agent Framework gives you both the staff and the pathway.
What it is
Agent Framework is a set of libraries for Python and .NET, developed in one public repository under the MIT licence. In Python it installs as agent-framework, with smaller packages such as agent-framework-core for those who want fewer dependencies.
| Part | What it does |
|---|---|
| Agent | A model-backed worker with instructions, tools and conversation state |
| Tools | Functions the agent can call, or tools discovered from MCP servers |
| Workflow | A typed graph of agents and functions with edges, branches and parallel paths |
| Orchestration patterns | Ready-made sequential, concurrent, handoff, group chat and Magentic patterns |
| Middleware | Interceptors around agent turns, tool calls and model calls |
| Checkpointing | Saving workflow state so long processes can pause and resume |
| Human-in-the-loop | Workflow steps that wait for a person's input or approval |
| Connectors | Adapters for Azure OpenAI, OpenAI, Microsoft Foundry and other providers |
| Telemetry | OpenTelemetry traces and metrics for agents and workflows |
| Declarative workflows | Workflows described in YAML and loaded at runtime |
Where it came from
Microsoft built two separate agent projects in the early 2020s. Semantic Kernel became the company's production SDK for AI features in C#, Python and Java, with plugins, connectors and planning. AutoGen came out of Microsoft Research and became well known for conversational multi-agent systems, including the Magentic-One team of agents for complex web and file tasks.
Developers had to choose between a stable enterprise SDK and an experimental multi-agent toolkit. On 1 October 2025 Microsoft announced Agent Framework in public preview as the single path forward, building on both. Semantic Kernel's kernel, plugin and connector ideas carried over. AutoGen's multi-agent ideas were rebuilt on a graph-based workflow engine. Microsoft published migration guides from both projects, and AutoGen moved into maintenance mode.
Agent Framework reached version 1.0 for both .NET and Python in April 2026, described by Microsoft as production-ready, with stabilised orchestration patterns, MCP support and OpenTelemetry integration.
How it works
Step 1: Create an agent
You choose a chat client for a model provider, then create an agent with a name, instructions and tools. In Python a function becomes a tool through a decorator; in .NET you register methods as functions. The agent can then be run with a message and returns a response.
Step 2: Keep conversation state
An agent keeps conversation history in a thread or session object that you pass between turns. Context providers can add information, such as memories or retrieved documents, before each model call. This is how you give an agent long-term memory or RAG.
Step 3: Add middleware
Middleware runs around three layers: the agent turn, each tool call and each model call. Typical uses are logging, blocking a tool, redacting personal data, checking content and enforcing policy. Because middleware is code, it applies on every call.
Step 4: Build a workflow
When one agent is not enough, you build a workflow. Nodes, called executors, are agents or plain functions. Edges connect them and can carry conditions. Messages flow along edges with types, so the framework can check that one step's output fits the next step's input.
For common shapes there are ready-made orchestrations:
- Sequential: each agent works on the output of the previous one.
- Concurrent: several agents work on the same input at once, and results are combined.
- Handoff: an agent passes control to a more suitable agent.
- Group chat: several agents take turns in a shared conversation, managed by a coordinator.
- Magentic: a manager agent plans, assigns tasks to specialists and tracks progress, based on the Magentic-One design.
Step 5: Pause, resume and ask people
Workflows can save checkpoints, so a process that waits hours for an event or a person can resume without starting again. A workflow step can request information from a human and wait for the reply, which is how approval gates are built.
Step 6: Observe and deploy
Agents and workflows emit OpenTelemetry data, so traces can go to Azure Monitor or any compatible backend; see OpenTelemetry. You can host agents in your own services, or use Microsoft's managed agent services on Azure.
Key concepts and terms
| Term | Meaning |
|---|---|
| Agent | A model-driven worker with instructions and tools |
| Chat client | The adapter that connects an agent to a model provider |
| Thread or session | Stored conversation state for an agent |
| Context provider | Component that adds memory or data before each model call |
| Middleware | Code that intercepts agent, tool or model calls |
| Workflow | A typed graph of executors connected by edges |
| Executor | A node in a workflow: an agent or a function |
| Edge | A connection that routes messages between executors |
| Checkpoint | Saved workflow state used to pause and resume |
| Magentic orchestration | A manager-led pattern where a planner assigns work to specialists |
| Semantic Kernel | Microsoft's earlier enterprise AI SDK, a predecessor |
| AutoGen | Microsoft Research's earlier multi-agent framework, a predecessor |
A worked example
This scenario is illustrative. An insurer builds a claims intake process.
- An "intake" agent reads an emailed claim, extracts fields such as policy number, date and amount, and returns them as a typed object.
- A function executor checks the policy in the core system. If the policy is not active, an edge routes the claim to a rejection letter step.
- For active policies, a concurrent step runs two agents at once: one checks for fraud signals, the other estimates repair cost from attached documents.
- If the estimate is above 5,000 euros or the fraud agent raises a flag, the workflow requests human review and saves a checkpoint. A claims handler may reply hours later; the workflow resumes from the checkpoint.
- Middleware removes national ID numbers from anything sent to the model and logs every tool call.
- Traces go to the insurer's monitoring system. In a pilot of 300 claims, the team uses traces to find that 22 claims stalled because the document agent could not read scanned images, and adds an OCR tool.
Strengths and limitations
Strengths:
- One framework for both simple agents and complex, typed multi-agent workflows.
- First-class .NET support as well as Python, which suits Microsoft-centred organisations.
- Built-in checkpointing and human-in-the-loop for long-running processes.
- Ready-made orchestration patterns, including Magentic.
- OpenTelemetry, middleware and Entra ID integration for enterprise needs.
- Migration paths from Semantic Kernel and AutoGen.
Limitations:
- It is newer than its predecessors, so examples and community answers are still catching up.
- Best integration is with Azure and Microsoft Foundry; other stacks need more setup.
- The workflow model has more concepts than a minimal SDK.
- Teams on Semantic Kernel or AutoGen must plan a migration.
- Some features may differ between the .NET and Python versions.
How it compares with alternatives
| Option | Strength | Trade-off |
|---|---|---|
| Microsoft Agent Framework | .NET and Python, typed workflows, enterprise controls | Best inside Microsoft's stack |
| LangGraph | Mature graph model and checkpoints in Python and JavaScript | No .NET version |
| Google ADK | Workflow agents, eval and deploy tools for Google Cloud | Best inside Google's stack |
| OpenAI Agents SDK | Very small set of primitives | Fewer workflow features |
| CrewAI | Role-based crews and flows | Python only |
Practical guidance and best practices
- Start with a single agent and move to a workflow only when you need fixed steps, branching or approvals.
- Use typed messages between workflow steps so errors appear early.
- Put deterministic checks, such as policy lookups, in function executors rather than agents.
- Add checkpoints wherever a workflow may wait on people or outside systems.
- Enforce data rules in middleware, not only in prompts.
- Send telemetry to your normal monitoring stack from day one.
- Gate payments, deletions and external messages behind human-in-the-loop steps.
- Keep a regression set of real cases and re-run it as evals after changes.
Common mistakes
- Using group chat for tasks that are really a simple sequence.
- Letting agents do deterministic work, such as arithmetic or lookups, that a function would do reliably.
- Building long workflows without checkpoints.
- Migrating from AutoGen by copying old patterns instead of using workflows.
- Sending unredacted personal data to models because middleware was never added.
Frequently asked questions
What is Microsoft Agent Framework?
It is Microsoft's open-source framework for building AI agents and multi-agent workflows in Python and .NET, combining ideas from Semantic Kernel and AutoGen.
Does it replace Semantic Kernel and AutoGen?
It is the recommended path for new work and the successor to both. Microsoft provides migration guides, and AutoGen is in maintenance mode.
When was it released?
It was announced in public preview on 1 October 2025 and reached version 1.0 for .NET and Python in April 2026.
Can it use non-Microsoft models?
Yes. Besides Azure OpenAI and Microsoft Foundry, it has connectors for OpenAI and several other providers, including local models.
What is Magentic orchestration?
It is a pattern in which a manager agent plans a task, assigns parts to specialist agents and tracks progress until done. It comes from Microsoft Research's Magentic-One system.
Does it support MCP?
Yes. Agents can discover and call tools from MCP servers, and the framework also works with the A2A protocol for agent-to-agent communication.
Glossary
- Agent: a model-driven worker with tools.
- Workflow: a graph of steps with typed messages.
- Executor: a workflow node.
- Edge: a link that carries messages between nodes.
- Middleware: code that intercepts calls.
- Checkpoint: saved state for pause and resume.
- Context provider: a source of extra context before model calls.
- Magentic: a manager-and-specialists orchestration.
- OpenTelemetry: an open standard for traces and metrics.
Summary
Microsoft Agent Framework unifies Semantic Kernel and AutoGen into one open-source framework for Python and .NET. It offers agents for model-driven work and typed, checkpointed workflows for structured processes, with ready-made orchestration patterns, middleware, telemetry and protocol support. It reached 1.0 in April 2026. Use plain functions for deterministic steps, workflows for anything with approvals or long waits, and middleware and telemetry from the start.
Related
It belongs to Agent SDKs in my stack, under Agentic AI. Others in the same group: