Microsoft Agent Framework
Build, orchestrate, and deploy production-grade AI agents and multi-agent workflows in Python and .NET.
Per-dimension scores and reasoning
Evidence shows: repository includes security policy (SECURITY.md) and CodeQL workflow, but lacks specific implementation details for least privilege, user confirmation, data flow transparency, sensitive data handling, dependency security, external effects, and rollback. Deductions: these aspects are only partially mentioned or not clearly specified. Source attribution: clearly Microsoft official organization, but no specific author or contributor information provided.
Evidence shows: README and code structure are consistent, but no specific implementation for dependency availability and failure messages. Deductions: self-consistency is adequate, but dependency availability and failure messages are only partially covered.
Evidence shows: README clearly defines target audience and scenarios, provides multiple examples, but does not clearly define capability boundaries and trigger precision. Deductions: environment fit is good, but capability boundaries and trigger precision are not fully explained.
Evidence shows: README structure is clear, provides installation instructions and examples, but does not provide naming stability, known limitations, and version changelog. Deductions: information architecture and installation notes are good, but naming stability, known limitations, and version changelog are missing.
Evidence shows: README provides quickstart and examples, but does not provide cost-benefit analysis. Deductions: output usability and marginal value are adequate, but cost-benefit is not explained.
Evidence shows: claims in README are not supported by specific evidence, and facts and inferences are not separated. Deductions: claim traceability, cross-source corroboration, and fact-inference separation are insufficient.
- Static review cannot verify actual runtime behavior; all scores are based on source code and documentation.
- Dependency security and external effects are not detailed in the provided files; further review is needed.
- Version changelog and known limitations are not mentioned in README, which may affect user decisions.
What does this agent do, and when should you use it?
Microsoft Agent Framework (MAF) is an open framework for building production-grade AI agents and multi-agent workflows in Python and C#/.NET. It supplies Agent/AIAgent interfaces, provider integrations, middleware, and graph-based patterns for sequential, concurrent, handoff, and group collaboration flows. Workflows include checkpointing, streaming, human-in-the-loop control, and time-travel, while OpenTelemetry integration supports distributed tracing, monitoring, and debugging. The repository also includes YAML declarative agents, Agent Skills, DevUI, and local and cloud hosting samples, including Foundry-hosted agents, A2A, Azure Functions, and Durable Task hosting.
In Python, a developer creates an Agent with a FoundryChatClient, name, and instructions, then calls await agent.run(...) to obtain a model response. In .NET, an AIProjectClient can create an AIAgent through AsAIAgent(...), which is invoked with await agent.RunAsync(...). Middleware processes requests, responses, exceptions, and custom pipelines, while graph-based workflows coordinate sequential, concurrent, handoff, or group-collaboration agent execution. The framework documents checkpointing, streaming, human-in-the-loop control, OpenTelemetry telemetry, YAML declarative agents, and samples for deployment to Foundry-hosted infrastructure.
- A Python team turning a prompt-driven prototype into a restartable, checkpointed multi-agent workflow.
- A C#/.NET service team integrating a Microsoft Foundry-backed AIAgent through AsAIAgent and RunAsync.
- An application developer coordinating multiple agents through sequential, concurrent, handoff, or group-collaboration patterns.
- A platform team that needs OpenTelemetry-based tracing, monitoring, and debugging for agent execution.
- An Azure team evaluating Foundry-hosted agents or the provided A2A, Azure Functions, and Durable Task hosting samples.
What are this agent's strengths and limitations?
- It supports both Python and C#/.NET and presents them as a consistent framework foundation.
- Its workflow model explicitly covers sequential, concurrent, handoff, and group collaboration patterns, plus checkpointing, streaming, human control, and time-travel.
- OpenTelemetry is built in for distributed tracing, monitoring, and debugging.
- It documents ecosystem support for Microsoft Foundry, Azure OpenAI, OpenAI, and the GitHub Copilot SDK, with local and cloud hosting samples.
- The Python quickstart depends on Azure CLI sign-in and Azure credentials; the .NET Foundry sample depends on an Azure project endpoint and Azure.Identity.
- Third-party servers, agents, code, and non-Azure Direct models are governed by their own terms, and users are responsible for related usage, costs, data flows, and compliance implications.
- The project places responsibility for application review, testing, content filters, metaprompts, and other responsible-AI mitigations on the application builder.
- Although several provider ecosystems are named, provider-specific setup is distributed across samples and environment-variable configuration must be checked for the chosen path.
How do you install or deploy this agent?
Python: run pip install agent-framework. The Python Foundry quickstart also requires az login and the azure.identity dependency used in the sample. Create Agent(client=FoundryChatClient(credential=AzureCliCredential()), name="HaikuAgent", instructions="..."), then call await agent.run("Write a haiku about Microsoft Agent Framework.").
.NET: run dotnet add package Microsoft.Agents.AI. For the README Foundry quickstart, also run dotnet add package Microsoft.Agents.AI.Foundry, dotnet add package Azure.AI.Projects, and dotnet add package Azure.Identity. Set AZURE_AI_PROJECT_ENDPOINT; AZURE_AI_MODEL_DEPLOYMENT_NAME is optional and the sample defaults it to gpt-5.4-mini.
How do you use this agent?
For a first Python invocation, construct Agent with FoundryChatClient(credential=AzureCliCredential()) and run print(await agent.run("Write a haiku about Microsoft Agent Framework.")) inside asyncio.run(main()). The endpoint, deployment name, and API version may be supplied through environment variables or passed directly to FoundryChatClient.
For .NET, read AZURE_AI_PROJECT_ENDPOINT, construct new AIProjectClient(new Uri(endpoint), new DefaultAzureCredential()).AsAIAgent(model: deploymentName, instructions: "...", name: "HaikuAgent"), then call Console.WriteLine(await agent.RunAsync("Write a haiku about Microsoft Agent Framework.")).
How does this agent compare with similar options?
The project provides migration guides from Semantic Kernel and AutoGen; the supplied material does not claim complete feature parity with either framework.