AIAI Agent Dev Labaiagentdev-lab.com
Sign in to save progressSign in
Menu

AB-620 · LESSON 2

Integrate and extend agents in Copilot Studio40–45% · 4 topics

Learn to connect enterprise knowledge, add tools, configure multi-agent collaboration and integrate with Azure.

2.1Connect to enterprise knowledge sources

Concept

Knowledge is the information an agent uses to ground its answers. You can connect SharePoint, OneDrive, public websites, uploaded files, Dataverse, Azure AI Search and Microsoft Graph connectors (such as ServiceNow). Sources such as SharePoint are searched based on the user's permissions.

How it works / steps

  1. Identify where the information needed for answers lives
  2. Add SharePoint sites and websites as knowledge
  3. Connect an Azure AI Search index for large or custom document sets
  4. Connect data from other systems with Graph connectors or connectors
  5. Write knowledge descriptions and verify answer quality with evaluation

Workplace example

An IT help agent uses SharePoint procedures and ServiceNow knowledge articles (via a Graph connector) as knowledge, answering only with information the user is allowed to view.

Common confusions and exam tips

Mini quiz

Q1. You want to bring knowledge articles stored in ServiceNow into Microsoft 365 and use them for the agent’s search and grounding. Which is most suitable?

  1. Computer use
  2. Power Platform connector
  3. HTTP request node
  4. Copilot connector
Show answer and explanation

Answer: D Copilot connector

Copilot connectors are suited to ingesting external content for reference. [Other options] Power Platform connectors and HTTP requests call actions on the spot; they do not ingest and index articles in Microsoft 365. Computer use automates UI operations.

Q2. You want the sales agent to register order data in SAP. Which option appropriately implements the register (write) operation?

  1. Add the Power Platform connector for SAP (or an existing custom connector) as a tool
  2. Ingest SAP order data into Microsoft 365 with a Copilot connector and have a generative answers node confirm the registration
  3. Add the SAP order screen manual to knowledge and have the agent guide users through input steps
  4. Save the order in a Dataverse table and have staff transfer it to SAP in a batch at night
Show answer and explanation

Answer: A Add the Power Platform connector for SAP (or an existing custom connector) as a tool

Operations such as writes are done with tools (connectors). Knowledge is for reference and cannot register data. [Other options] Copilot connectors ingest for search and cannot write. Manual transfer or step guidance does not mean the agent registered the order.

Reference links

2.2Add tools to agents

Concept

Tools are how an agent performs actions in external systems. You can add Power Platform connectors (prebuilt and custom), agent flows, REST APIs (OpenAPI), MCP servers, prompts (AI actions) and computer use (UI automation).

How it works / steps

  1. Identify the actions needed (look up, register, send)
  2. Use a prebuilt connector if one exists; otherwise use a custom connector, OpenAPI or MCP
  3. Write tool names, descriptions and input descriptions so the model can choose them easily
  4. Decide authentication (the user's credentials or the maker's credentials)
  5. Consider computer use for legacy systems without an API

Workplace example

Use an ERP custom connector for inventory lookup, an MCP server for knowledge base search, and computer use to operate an old ordering screen that has no API.

Common confusions and exam tips

Mini quiz

Q1. You want to enter data through the UI of an old internal business app that has no API. Which capability is appropriate?

  1. Custom connector
  2. HTTP request node
  3. Copilot connector
  4. Computer use
Show answer and explanation

Answer: D Computer use

When there is no API, computer use, which automates UI operations, is an option. [Other options] Custom connectors and HTTP requests require an API. Copilot connectors ingest data for search and cannot operate a UI.

Q2. You want the agent to use the tools of an MCP server that is already published internally. What is the appropriate action?

  1. Register the MCP server as a connected agent and delegate work with A2A
  2. Add the MCP server’s tool definitions to knowledge and reference them with a generative answers node
  3. Connect to the MCP server in Copilot Studio and add the tools it provides to the agent
  4. Create a topic for each tool that calls it individually with an HTTP request node
Show answer and explanation

Answer: C Connect to the MCP server in Copilot Studio and add the tools it provides to the agent

Connecting the MCP server lets you use the tools defined on the server as they are. [Other options] Reimplementing with HTTP requests is unnecessary effort. A2A is for agent-to-agent collaboration, not for connecting MCP servers. Tools cannot be called from knowledge.

Reference links

2.3Configure multi-agent collaboration from Copilot Studio

Concept

In a multi-agent setup, the main agent delegates work to child agents (components inside the same agent) or connected agents (separately published Copilot Studio agents, Foundry agents, Fabric data agents, A2A agents and more). Separating roles makes maintenance and reuse easier.

How it works / steps

  1. Split the business by role and define each agent's responsibilities
  2. Use child agents for small, tightly coupled parts and connected agents for parts managed and reused independently
  3. Write each agent's description so the main agent can call it appropriately
  4. Configure how much conversation history to pass
  5. Verify the call path in testing

Workplace example

An employee front-desk agent delegates HR questions to the HR department's published agent and data aggregation to a Fabric data agent.

Common confusions and exam tips

Mini quiz

Q1. Which statement correctly describes the difference between MCP and the A2A (Agent2Agent) protocol?

  1. MCP is only for integration between Microsoft products, and A2A is only for integration with agents in other clouds
  2. MCP is for agents to connect to tools and data; A2A is for agents to collaborate with each other
  3. MCP is for agents to collaborate with each other; A2A is for agents to connect to tools and data
  4. MCP is a standard for knowledge search, and A2A is a standard for delegating agent authentication and permissions
Show answer and explanation

Answer: B MCP is for agents to connect to tools and data; A2A is for agents to collaborate with each other

Their roles differ: tool connection is MCP, and collaboration between agents is A2A. [Other options] The third option reverses the roles. Neither MCP nor A2A is limited to a particular vendor. A2A is not a standard for delegating authentication.

Q2. You want to connect a quote creation agent with a credit check agent. Which multi-agent design is appropriate?

  1. Combine both agents’ instructions and operate them as a single agent
  2. Have users ask the same question to the quote agent and the credit agent in turn
  3. Have a parent agent delegate work to specialist agents according to their roles
  4. Regularly ingest the quote agent’s answers into the credit agent’s knowledge
Show answer and explanation

Answer: C Have a parent agent delegate work to specialist agents according to their roles

Splitting roles and delegating from a parent agent is better for maintainability and quality. [Other options] Combining them makes role-specific maintenance and evaluation hard. Knowledge cannot connect them in real time, and asking users to enter things twice is not integration.

Reference links

2.4Integrate agents with Azure

Concept

Copilot Studio agents can be extended with Azure capabilities. Use Azure AI Search as knowledge, Foundry models (custom or fine-tuned) in prompts, Foundry agents as connected agents, and Azure Functions or API Management as the foundation for tools.

How it works / steps

  1. Identify requirements Copilot Studio alone cannot meet (custom models, large-scale search, complex processing)
  2. Connect an Azure AI Search index as knowledge
  3. Use models deployed in Foundry
  4. Call Foundry Agent Service agents as connected agents
  5. Turn Azure Functions APIs into tools with OpenAPI or custom connectors

Workplace example

Ingest millions of technical documents into Azure AI Search and connect it as knowledge, and perform specialized classification with a fine-tuned model in Foundry.

Common confusions and exam tips

Mini quiz

Q1. Tens of thousands of internal manuals are already indexed in Azure AI Search. What is the quickest configuration to ground answers in them?

  1. Fetch Azure AI Search results with an HTTP request node every time and format them
  2. Connect Azure AI Search as a knowledge source
  3. Upload the manuals to Copilot Studio as files
  4. Move the manuals to SharePoint and add the site as knowledge
Show answer and explanation

Answer: B Connect Azure AI Search as a knowledge source

Using the existing index avoids re-ingestion and enables features such as vector search. [Other options] Migrating or re-uploading wastes the existing index. Building it yourself with HTTP requests is not the quickest.

Q2. You want a Copilot Studio agent to use an advanced document analysis agent already built in Microsoft Foundry. Which method is appropriate?

  1. Integrate the Foundry agent into Copilot Studio and call it as a connected agent
  2. Port the Foundry agent’s instructions and tool definitions to the Copilot Studio agent
  3. Select the model used in Foundry in a Copilot Studio custom prompt as a substitute
  4. Save the Foundry agent’s analysis results to SharePoint and reference them as knowledge
Show answer and explanation

Answer: A Integrate the Foundry agent into Copilot Studio and call it as a connected agent

You can integrate and call Foundry agents from Copilot Studio. Duplicating it means managing two copies. [Other options] Porting or selecting the model alone cannot reproduce the tools and processing on the Foundry side. Referencing saved results is not calling the agent.

Reference links

Lessons are written by this site based on public information and do not guarantee exam content. Product specifications change, so check the latest information at the reference links.

Notes (unofficial, original questions, no pass guarantee)