Build Persistent Agents on Azure AI

Covers building persistent agents on Azure AI Foundry using the Microsoft Agent Framework Python SDK. The architecture runs a user query through AzureAIAgentsProvider to the Azure AI Agent Service (a persistent, server-hosted agent), executed via agent.run() or agent.run_stream(), with three tool categories - plain functions, hosted tools (code interpreter, file search, web search), and MCP tools - and conversation state kept in an AgentThread.

Installed via pip install agent-framework --pre (or the Azure-only agent-framework-azure-ai package), configured through AZURE_AI_PROJECT_ENDPOINT, AZURE_AI_MODEL_DEPLOYMENT_NAME, and BING_CONNECTION_ID (for web search), and authenticated with AzureCliCredential for development or DefaultAzureCredential for production.

async with (
    AzureCliCredential() as credential,
    AzureAIAgentsProvider(credential=credential) as provider,
):
    agent = await provider.create_agent(name="MyAgent", instructions="You are a helpful assistant.")
    result = await agent.run("Hello!")

It walks through progressively richer patterns: a basic agent with just instructions; a function-tool agent where plain Python functions (with Annotated[type, Field(description=...)] parameter typing) are passed directly to tools= and auto-converted to callable tools; a hosted-tools agent combining HostedCodeInterpreterTool and HostedWebSearchTool for code execution and Bing search; streaming responses via agent.run_stream(), printing text chunks as they arrive; multi-turn conversation threads via agent.get_new_thread(), where a second agent.run() call on the same thread retains context from the first, with the thread's conversation_id saved for later resumption; and structured outputs, where a Pydantic model with extra="forbid" is passed as response_format and the result is parsed back with model_validate_json.

Three provider methods round out the surface: create_agent() creates a new agent on the Azure AI service, get_agent(agent_id) retrieves an existing one by ID, and as_agent(sdk_agent) wraps an existing SDK Agent object without an HTTP call. Five hosted/MCP tool types are documented: HostedCodeInterpreterTool (execute Python), HostedFileSearchTool (search vector stores), HostedWebSearchTool (Bing search), HostedMCPTool (service-managed MCP), and MCPStreamableHTTPTool (client-managed MCP, connected via URL).

A complete example combines all of this into a "ResearchAssistant" agent wired with a weather function, code interpreter, web search, and a client-managed MCP tool pointed at Microsoft's own documentation MCP endpoint, running non-streaming, streaming, and structured-output calls on the same conversation thread.

Use this reference when building Python agents on Azure AI Foundry that need persistent, server-hosted execution, multi-turn conversation state, and a mix of custom function tools, hosted tools (code execution, file/web search), and MCP integrations. It is specific to the Azure-hosted agent path of the Microsoft Agent Framework - not other framework backends or non-Azure LLM providers.

Input is agent instructions, tool definitions (functions, hosted tools, or MCP tools), and user messages, optionally scoped to a conversation thread. Output is agent text responses, streamed text chunks, or structured Pydantic-validated objects when a response_format is specified.

Built on agent-framework-azure-ai, Azure AI Agent Service, azure.identity for authentication (AzureCliCredential/DefaultAzureCredential), Bing for web search, and the Model Context Protocol for both service-managed (HostedMCPTool) and client-managed (MCPStreamableHTTPTool) tool integration.

Python developers building persistent, tool-using agents on Azure AI Foundry who need function calling, hosted code/search tools, MCP integration, multi-turn threads, or structured output in one framework.