Integrate and extend agents in Copilot Studio Study Guide — AB-620

Integrate and Extend Agents in Copilot Studio (40%)

This domain tests your ability to connect Copilot Studio agents to external systems, data sources, and other agents. Expect scenario-based questions that ask you to choose the right integration mechanism for a given constraint—authentication requirement, latency tolerance, data source type, or orchestration pattern. Knowing when to use each approach matters more than memorizing UI steps.

Configuring Tools and Actions

Agents invoke external logic through actions, which fall into three categories:

• Power Automate flows – Best when you need multi-step logic, approvals, or connectors that don't have a direct Copilot Studio action. Flows are called synchronously; the agent waits for a response. Keep flow execution under the 120-second timeout or the action will fail silently from the agent's perspective.
• Connector actions – Pre-built actions surfaced directly in Copilot Studio from Power Platform connectors. Faster to configure than flows but less flexible for branching logic.
• REST API (HTTP request action) – Use when calling an API that has no connector. Requires manual input/output schema definition. Supports JSON bodies; always define the response schema so the agent can parse outputs.

Trap: Power Automate flows must be in the same environment as the agent. Cross-environment flow calls are not supported.

Integrating Agents with Microsoft Foundry

Azure AI Foundry integration allows agents to leverage custom AI models, prompt flows, and fine-tuned endpoints hosted in Foundry. Key points:

• Connect via the AI Foundry tool type in Copilot Studio; you supply the endpoint URL and key.
• Use Foundry when the built-in generative AI in Copilot Studio is insufficient—e.g., domain-specific models or structured reasoning chains.
• Foundry connections authenticate via API key or managed identity; prefer managed identity in production to avoid secret rotation issues.
• Output from Foundry nodes must be mapped back into the conversation context explicitly.

Integrating Agents with Model Context Protocol (MCP) Servers

MCP is an open protocol for exposing tool capabilities to AI agents in a standardized way.

• In Copilot Studio, add an MCP server as an external tool source; the agent discovers available tools automatically from the server's manifest.
• MCP servers can be local (SSE transport) or remote (HTTP/SSE). Remote servers require a publicly reachable endpoint.
• The agent uses the tool descriptions in the MCP manifest to decide when to invoke a tool—write clear, specific tool descriptions to prevent misrouting.
• Trap: MCP integration does not replace connector actions for licensed Microsoft services. Use MCP for custom or third-party tooling.

Custom Connectors and OpenAPI/Postman Imports

• Import an OpenAPI (Swagger) definition to auto-generate a connector; Copilot Studio can then surface that connector's operations as actions.
• Postman collections can be exported as OpenAPI 3.0 and imported the same way—validate the export first, as Postman sometimes produces non-compliant schemas.
• Custom connectors support OAuth 2.0, API key, and basic auth. The connector's auth model must match the target API exactly.
• Limit: custom connector actions must complete within 120 seconds; long-running operations need a polling or webhook pattern.
• After import, manually review and annotate operation summaries and descriptions—the agent's tool-selection logic depends on these descriptions.

Integrating Agents with Microsoft Fabric