Skip to lesson

MicrosoftMicrosoft Certified: Multi-Agent AI Solutions Expert2026-102.4 Implement multi-agent orchestration

Module 02 - Specialized Sub-Agent Tools

Completed by opening the official lab

Why it matters

Splitting one large prompt into sub-agents is only useful when each one gets its own instructions, its own tools, and a typed contract with the supervisor.

What to practise

Extend the travel supervisor from Module 01, a LangGraph create_react_agent that calls MCP tools. Add two @tool sub-agents with Pydantic input schemas: find_places, a one-shot selector forced to one call with bind_tools(..., tool_choice="required"), and an itinerary ReAct agent that writes trips to Azure Cosmos DB through MCP tools. Bind the supervisor with parallel_tool_calls=True, and add the hybrid search and trip tools to the MCP server. Watch the identity wrapper that stops the sub-agent from guessing the user and tenant IDs and saving trips under the wrong partition key.

How it fits

The agent-as-tool pattern is 2.4.5’s bounded delegation, and the separate tool subsets apply 2.4.3’s tool-access rules. The supervisor’s in-memory checkpointer is not the durable state 2.4.12 asks for; note what would replace it.