microsoft-agent-framework

Microsoft Agent Framework

Overview

Use this skill to implement or explain Microsoft Agent Framework usage in Python. Prefer Microsoft Learn docs for conceptual guidance and use Context7 to fetch exact snippets (package extras, Azure/OpenAI client specifics).

Workflow

1. Identify runtime and provider

• Confirm Python.
• Pick provider: OpenAI, Azure OpenAI, or Azure AI Foundry.
• Confirm required environment variables before coding.

1. Install and configure

• Use pip packages and extras for the provider you need.
• Load env vars from the shell or a .env file.

1. Create a basic agent

• Choose an agent type: ChatAgent, OpenAIResponsesClient, AzureOpenAIResponsesClient, or AzureAIAgentClient (Azure AI).
• Use OpenAIChatClient (OpenAI) or AzureOpenAIResponsesClient (Azure OpenAI) for common setups.
• Start with non-streaming, then add streaming if needed.

1. Add tools and functions

• Python: pass callables via tools=[...] on ChatAgent or per request.
• Use HostedCodeInterpreterTool when you need sandboxed Python execution.
• Use @ai_function(approval_mode="always_require") for human approvals and handle user_input_requests.

1. Orchestrate multi-agent workflows

• Use WorkflowBuilder and edges for simple graphs.
• Use fan-out/fan-in and branching edge groups when you need concurrency or routing.
• Use SequentialBuilder for pipeline workflows and workflow.as_agent() when you need a workflow to behave like a single agent.
• Use MagenticBuilder for manager/participant orchestration (advanced).
• Inspect AgentRunEvent outputs to debug.

1. Integrate external tools via MCP

• Use HostedMCPTool for Microsoft Learn MCP.
• Use MCPStreamableHTTPTool for HTTP/SSE MCP servers.

1. Add memory and storage

• Serialize/deserialize threads for persistence.
• Use a memory provider or chat message store for long-term history.

1. Add middleware

• Use agent-level middleware for cross-cutting concerns (logging, security).
• Add run-level middleware when behavior is per-request.