AB-620 Designing and Building Integrated AI Agent Solutions in Copilot Studio Questions and Answers

Authenticated governed access → Select service-principal authentication; Required curated vector index → Enter the Azure AI Search index name; Correct Azure AI Search service instance → Provide the Azure AI Search endpoint URL.

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: An Azure AI Search connection needs three distinct pieces of configuration. The endpoint identifies the search-service instance that hosts Fabrikam ' s indexed documents. The index name then selects the specific vector index containing the curated policies and procedures. Authentication determines how Copilot Studio is authorized to query that service. Because the scenario requires governed application access rather than a maker ' s personal identity, the service-principal option is the appropriate selection from the choices provided. Enabling citations is useful after retrieval is configured, but it does not establish the service connection or select an index. Microsoft recommends creating Azure AI Search through the supported data-source connection experience rather than manually assembling an unsupported endpoint/key combination. Administrators should grant the service principal only the permissions needed for query operations, protect its credentials, and maintain separate configuration per environment. Testing should confirm that the intended index—not another index in the same service—is queried, that unauthorized content is excluded, and that policy responses cite material that actually supports the generated guidance. Study Guide alignment: Integrate and extend agents in Copilot Studio > Connect to enterprise knowledge sources > Connect to Azure AI Search.

===============

Policy questions → Delegate to a knowledge-focused agent; Metric requirements → Delegate to a data-access agent; Transaction requirements → Delegate to a tools-first agent.

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: The architecture should separate responsibilities according to the kind of work each specialist performs. Policy and procedure questions depend on retrieval from curated, indexed content and therefore belong with a knowledge-focused agent. Operational measures such as backlog, SLA exposure, and parts shortages require governed, current business data; a data-access specialist, represented in this case by the Fabric Data Agent, is the correct destination. Creating a parts request or changing a service-case status is different from answering a question: it is an authenticated transaction and should be delegated to a tools-first agent that can invoke approved MCP operations. Putting all three responsibilities into one agent would increase the number of competing tools and knowledge sources, weaken routing precision, and duplicate logic already owned by specialist teams. The selected pattern also matches Fabrikam ' s requirement that capabilities evolve independently. Clear names, descriptions, inputs, and outputs remain essential because generative orchestration uses that metadata to decide which connected agent to invoke. This division gives the primary agent a coordination role while each specialist retains its own knowledge, privileges, release lifecycle, and audit boundary. Study Guide alignment: Integrate and extend agents in Copilot Studio > Configure multi-agent collaboration from Copilot Studio > Design multi-agent solutions in Copilot Studio.

===============

Callable external summarization agent → Add an external agent of type Microsoft Foundry; Correct Foundry project and agent → Enter the Foundry project endpoint URL to create the connection.

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: The existing specialist should be connected as a Microsoft Foundry agent, not converted into indexed knowledge. A connected agent remains callable and can execute its own summarization and classification behavior when Operations Concierge delegates a matching request. During connection creation, Copilot Studio uses the Microsoft Foundry project endpoint URL to identify the project context and make the relevant agent available for selection. This is distinct from a model endpoint used directly by a custom prompt and from a knowledge-source connection used only for retrieval. Publishing the Copilot Studio agent is part of making a completed design available to users, but it is not the action that establishes the Foundry connection. Adding the Foundry output to an index would turn prior output into retrievable content and would not invoke the specialist on demand. The connected agent should receive a precise description covering long-form transcript summarization and classification, and the integration should be tested with representative inputs to confirm that delegation occurs, the correct Foundry agent runs, and the returned result is incorporated into the main conversation. Study Guide alignment: Integrate and extend agents in Copilot Studio > Configure multi-agent collaboration from Copilot Studio > Integrate a Foundry agent.

===============

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: Fabrikam needs both reusable prompt behavior and explicit model governance. A reusable custom prompt template provides the stable summarization instructions, expected structure, and input/output contract that can be applied wherever summarization is required. Within that prompt ' s configuration, selecting the approved model from the Microsoft Foundry model catalog controls which model generates the output. These are separate controls: the template governs the task, while the model selection satisfies the approved-model requirement. Publishing only deploys the configuration already selected; it does not automatically choose or enforce an approved model. Citations are valuable for grounded policy answers but do not determine the model used for summarization. Azure AI Search is the retrieval system for indexed enterprise policy content and is not required merely to apply a long-form summarization template. The prompt should be tested with representative transcripts for completeness, classification accuracy, output schema, safety, and latency. Where Fabrikam already uses a connected Foundry agent for this specialist function, the prompt and model settings should remain owned by that component rather than being duplicated in the primary agent. Study Guide alignment: Integrate and extend agents in Copilot Studio > Integrate agents with Azure > Configure custom prompts to use the Foundry model catalog.

===============

Prevent most builders from modifying shared assets → Share the component collection while controlling edit access; Reuse components without copying → Create a component collection and move the shared components into it; Limit use to the intended primary agent → Restrict the component collection to the primary agent.

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: A component collection is the correct reusable unit for the escalation topics, MCP definitions, and safety disclaimer. Moving those assets into one collection creates a maintained source that multiple agents can consume without copying and later diverging. Sharing the collection with teammates while controlling edit access separates use permission from modification permission, allowing agent authors to reference the assets while reserving changes for platform engineering. If a sensitive or tightly coupled collection must be usable only by one designated primary agent, the collection can be restricted accordingly. Exporting the primary agent and copying its components would recreate the duplication problem that Fabrikam is trying to eliminate, while restricting every collection to the primary agent would prevent the required cross-agent reuse. The collection should be versioned and tested after changes because a shared update can affect several consumers. Ownership, review, and release responsibilities should also be documented so that reuse does not become an uncontrolled dependency. This design provides consistency, centralized governance, and a smaller maintenance surface across the three agents. Study Guide alignment: Plan and configure agent solutions > Plan an agent solution > Plan reusable agent components.

===============

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: The partner capability is described as an agent available only through a standardized agent-to-agent endpoint. That is the defining use case for an Agent2Agent connection. In Copilot Studio, the primary agent is configured with the external agent ' s A2A message endpoint and the applicable authentication details, after which orchestration can delegate shipment-tracking requests to it. An MCP server exposes tools or resources rather than an independently orchestrated partner agent, so option C changes the integration contract. An indexed knowledge source is also unsuitable because shipment status is dynamic and the requirement is delegation, not retrieval from a copied corpus. Generative responses in the primary agent would have no authoritative live shipment source and would violate the defined architecture. The A2A description should identify shipment tracking precisely so that it does not compete with unrelated agents or tools. Because external-agent support can be preview functionality, Fabrikam should also validate regional availability, authentication, error behavior, latency, and governance before production use. Study Guide alignment: Integrate and extend agents in Copilot Studio > Configure multi-agent collaboration from Copilot Studio > Create a multi-agent solution by using A2A protocol.

===============

A. Use a Microsoft Graph connector to index the partner’s travel advisory content. Microsoft 365 Copilot connectors—formerly called Microsoft Graph connectors—index non-Microsoft enterprise content into Microsoft Graph, making it available for semantic knowledge grounding and citations. Microsoft Learn: Copilot connectors

D. Add the internal policy documents as a knowledge source. The SharePoint PDFs and policy documents should be connected as agent knowledge so generative answers can retrieve and summarize the relevant content. This is scalable and avoids manually maintaining document-specific topics. Microsoft Learn: SharePoint knowledge sources

B violates the requirement to avoid manually authored Q & A topics. C could produce ungrounded responses from unapproved public sources and would not satisfy the airline’s accuracy, security, or compliance requirements.

Requirement

Implementation method

Let the agent decide when to run a tool based on user input.

Enable generative orchestration.

Run a specific agent flow at a defined step in a topic.

Add a tool node.

Use a connector action to retrieve data during a conversation.

Add a tool node.

Comprehensive and Detailed Explanation From [Microsoft AB-620 Study Guide]

Generative orchestration is appropriate when the agent must independently determine whether a tool is relevant to the user’s request. Copilot Studio evaluates the user’s message, conversation context, and the tool’s name, description, inputs, and outputs before deciding whether to invoke it. Microsoft documentation confirms that generative orchestration allows an agent to choose tools based on their descriptions and the current user query.

When an operation must occur at a predetermined point in an authored topic, the operation should instead be inserted explicitly by using an Add a tool node. Selecting an agent flow from that node ensures that the flow executes at that exact position in the topic rather than being selected autonomously by the orchestrator.

The same node type is used when calling a connector action. From the topic canvas, the author selects Add a tool , chooses Connector , and then selects and configures the required connector operation. Adaptive Cards present interactive information, while Question nodes collect user input; neither directly invokes flows or connectors.

Relevant study-guide area: Plan and configure agent solutions → Configure topics → Add agent flows to a topic and Add tools to a topic . See Orchestrate agent behavior with generative AI , Add tools to custom agents , and Use connectors in Copilot Studio agents .

Outcome of individual evaluation cases → Pass or fail status

Behavior observed across repeated runs → Consistent outcomes

Details presented in evaluation output → Expected and generated responses

A recurring failure across several runs indicates a consistent pattern, not an intermittent result. However, evaluation results alone do not automatically establish the root cause or grant production approval; the failed cases still require investigation.

Current Dataverse balance → Connect to a Power Platform connector; Real-time Dynamics 365 shipping status → Connect to a Power Platform connector; Static PDF policy content → Upload the file to the agent ' s knowledge; Welcome message independent of enterprise data → Create a manual response in the topic.

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: The retrieval pattern should match the volatility and purpose of each source. Customer balances and shipping status are operational values that can change during the day, so the agent should query the systems through Power Platform connectors at runtime rather than rely on a static copy. The connector connection also preserves the applicable authentication and source-system controls. A fixed PDF policy is unstructured reference material; uploading it as agent knowledge allows generative retrieval over its contents without building a transactional integration. The welcome message is deterministic presentation text and does not require any enterprise lookup, so a manual message in the topic is simpler and more reliable than a generative call. A generative answer without enterprise data would be inappropriate for balances or shipping status because it has no authoritative current value. This separation limits latency and security exposure, keeps dynamic information live, and uses knowledge grounding only where retrieval is actually needed. Each connector should be described precisely and tested under the intended user identity to ensure the returned records respect access permissions. Study Guide alignment: Integrate and extend agents in Copilot Studio > Connect to enterprise knowledge sources; Add tools to agents.

===============

Authentication → Add the required authentication headers; Retrieve data → Set the HTTP method to GET; Structured reusable response → Select a JSON/Record response type, define the schema, and save the response to a topic variable.

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: The Send HTTP request node must express the API contract directly. If the service expects a token or key in a header, the required header is added using a protected variable or connection-derived value rather than hard-coded secret text. GET is the correct method for retrieving status without changing the external resource; PATCH would imply a partial update. Selecting a structured JSON or Record response type and defining its schema allows Copilot Studio to type the returned fields. Saving that response to a topic variable makes values available for conditions, formatting, adaptive cards, or later tool calls. The schema must reflect optional fields, arrays, and nested records accurately; otherwise valid responses can be rejected or fields can appear empty. The topic also needs branches for non-success status codes, timeouts, invalid JSON, and authentication failure. Where a reusable connector or REST tool already exists, that managed abstraction may be preferable to embedding HTTP details in multiple topics, but the node configuration shown is correct for the stated direct-request requirement. Study Guide alignment: Plan and configure agent solutions > Configure topics > Configure advanced agent responses with API and Send HTTP requests.

===============

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: An external agent that supports Agent2Agent is connected by supplying its A2A endpoint and the required authentication configuration. The endpoint represents a collaboration interface between independently orchestrated agents, enabling the Copilot Studio agent to delegate a task and receive the partner agent ' s response. Generative answers cannot call another agent merely because both use generative AI; they retrieve and synthesize from knowledge sources. A Power Platform connector could wrap an ordinary API, but it would not use the specified A2A agent contract and would lose agent-to-agent semantics. Adding the target as indexed knowledge would only make a corpus searchable and would not execute the target agent. The endpoint must be the A2A message endpoint, not just a website or agent-card URL. The builder should give the connection a discriminating description, configure authentication if the endpoint is protected, and validate handoff behavior in the test pane. Production planning should include timeout, failure, logging, and data- sharing boundaries between the two agents. Study Guide alignment: Integrate and extend agents in Copilot Studio > Configure multi-agent collaboration from Copilot Studio > Create a multi-agent solution by using A2A protocol.

===============

Approval stage → Wait for a response and branch according to the decision; Information-collection stage → Wait for a response and capture the submitted inputs.

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: An approval is a decision gate. The flow must pause until the reviewer responds, then branch to the appropriate approved, rejected, or other outcome path. Information collection also pauses, but its purpose is to capture supplied values and make them available to later actions in the same run. Sending a request or notification and continuing immediately would break both requirements because downstream processing could occur before the human response exists. The action should have a defined timeout and an escalation or cancellation path so a missing response does not leave the process suspended indefinitely. The flow should store the response, responder identity, timestamp, and any comments needed for audit. Sensitive information should not be requested through an inappropriate channel and should be protected in logs and variables. Testing must cover approval, rejection, corrected information, timeout, and unauthorized-responder cases. This produces a controlled human-in-the-loop design rather than a simple alert that has no effect on execution. Study Guide alignment: Plan and configure agent solutions > Create and monitor agent flows in Copilot Studio > Create a human-in-the-loop agent flow.

===============

Based on Microsoft’s documented dependencies, the technically complete sequence is:

Publish the agent.

Configure and enable the agent’s target integration channel.

Retrieve current agent configuration values.

Implement the Microsoft 365 Agents SDK in the app using the agent configuration.

Validate the end-to-end conversation workflow in a test environment.

Deploy the web app to production.

Microsoft states that an agent must be published before channels are configured. The Web app or Native app channel then provides the Agents SDK connection string, after which the SDK can be implemented and tested. Microsoft Learn

===============

Block unsafe content → Configure the content-filtering/moderation level for AI-generated content; Diagnose why a response was filtered → Review Azure Application Insights telemetry; Identify the triggering user input → Review conversation transcripts.

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: Content moderation is the preventive control that evaluates generative input and output for harmful categories and blocks responses that violate the configured threshold. A stricter level reduces the likelihood that unsafe material is returned, although it may also increase false positives and reduce answer coverage. Application Insights provides operational telemetry for investigating filtered or failed behavior when appropriate logging is enabled, including events and message details subject to the organization ' s privacy settings. Conversation transcripts are the direct source for identifying which user utterance preceded the filter decision; connector run history would only be relevant if a connector failed and cannot reveal the original conversational input in this scenario. Logging must be configured carefully because prompts and responses can contain personal or sensitive information. Access to telemetry and transcripts should be restricted, retention should follow policy, and sensitive activity properties should not be enabled casually. The team should test both user prompts and model-generated responses, document expected safety behavior, and preserve a human review and escalation path for legitimate requests that are incorrectly blocked. Study Guide alignment: Plan and configure agent solutions > Plan an agent solution > Plan responsible AI strategy; Evaluate security and governance considerations.

===============

Capture response → Use the action ' s response outputs in later steps; Continue the same run → Allow the action to wait for and return a response before continuing; Require a manual decision → Add a human approval action.

Comprehensive and Detailed Explanation From Microsoft AB-620 Study Guide: A human-in-the-loop flow must preserve execution state while it waits. The approval or information action pauses the current run and returns a response when the designated person acts; triggering a new flow would lose that direct continuation contract. The response outputs—decision, comments, supplied fields, responder, and timestamps as applicable—are then referenced by downstream conditions and actions. A human approval action provides the explicit decision gate required before processing continues. Sending a notification or message alone does not wait and cannot guarantee that the decision controls the flow. The designer should configure expiration, escalation, reassignment, and rejection paths so the run cannot remain suspended without governance. Responses should be validated before use, and sensitive values should be protected in variables and history. Test approval, rejection, timeout, duplicate response, and unauthorized response scenarios. These controls ensure that human oversight changes execution deterministically instead of acting as an informational side channel. Study Guide alignment: Plan and configure agent solutions > Create and monitor agent flows in Copilot Studio > Create a human-in-the-loop agent flow.

===============