Build Persistent Agents on Azure AI

This skill teaches an agent to build persistent agents on Azure AI Foundry using the Microsoft Agent Framework Python SDK. It wires a query through AzureAIAgentsProvider to the Azure AI Agent Service, which persists the agent across calls; agents run via agent.run() or stream via agent.run_stream(), and can call function tools, hosted tools (code interpreter, file search, web search), or MCP tools. Conversation state persists in an AgentThread, and threads can be resumed later by their conversation_id. The provider itself exposes create_agent() to register a new agent on the Azure AI service, get_agent(agent_id) to retrieve one that already exists, and as_agent(sdk_agent) to wrap an existing SDK agent object without an extra HTTP call.

Use it when writing Python code that needs an agent hosted and persisted by Azure AI Foundry, rather than a locally run agent loop - for example, an assistant that must keep its identity and tool set across multiple conversation turns and sessions. It needs an Azure AI Foundry project endpoint, a model deployment, and Azure credentials (AzureCliCredential for development, DefaultAzureCredential for production) already set up; it is not a fit for work that does not touch the Azure AI Agent Service.

Inputs: an Azure AI project endpoint and model deployment name as environment variables (AZURE_AI_PROJECT_ENDPOINT, AZURE_AI_MODEL_DEPLOYMENT_NAME), an Azure credential, an agent name and instructions, an optional list of tools (plain Python functions, HostedCodeInterpreterTool, HostedFileSearchTool, HostedWebSearchTool, HostedMCPTool, or MCPStreamableHTTPTool), and an optional response_format Pydantic model for structured output. Outputs: a result from agent.run() / agent.run_stream() exposing .text (or, with response_format set, JSON validated into the given Pydantic model), plus an AgentThread whose conversation_id can be stored to resume the conversation later. The skill's conventions call for entering the credential and provider as async context managers, passing plain functions directly to tools= (they are auto-converted to an AIFunction), and describing function parameters with Annotated[type, Field(description=...)] so the model knows what each argument means. It also ships reference files covering hosted-tool patterns, MCP integration, thread management, and advanced topics like OpenAPI tools and citations.

Azure AI Agent Service through AzureAIAgentsProvider, Azure Identity (azure.identity.aio), Bing web search through HostedWebSearchTool (needs a BING_CONNECTION_ID), and MCP servers through HostedMCPTool (service-managed, preferred when the MCP server should be managed by the Azure service) or MCPStreamableHTTPTool (client-managed, for example against https://learn.microsoft.com/api/mcp).

Python developers building agents on Azure AI Foundry who need persistence, hosted or MCP tools, streaming, and structured output rather than a self-managed agent loop.

pip install agent-framework --pre