AB-620 Practice Questions

AI agents can automate tasks on user devices by leveraging Robotic Process Automation (RPA). In the Microsoft ecosystem, this is achieved by integrating an agent, like a Microsoft Copilot Studio copilot, with Power Automate for desktop. This integration allows the agent to trigger desktop flows that interact with applications and websites directly on a user's machine. This functionality effectively gives the agent a "computer use capability," extending its skills from conversational AI to performing hands-on digital tasks.

Static responses: These are predefined conversational replies and do not involve executing actions on a device.

Variables: These are used to store and manage data within the agent's logic, not to perform external tasks.

Logs: These are records of an agent's operations used for debugging and analysis, not for task automation.

Microsoft Learn, "Use Power Automate flows in Microsoft Copilot Studio". This document explains how to call cloud and desktop flows from a copilot. See the section: "Call a desktop flow".

Microsoft Learn, "Overview of Power Automate for desktop". This document describes how Power Automate for desktop enables UI-based automation (RPA). See the section: "Why use Power Automate for desktop?".

Adaptive Cards are a platform-agnostic method for sharing and displaying UI content in a common and consistent way. Within Microsoft Copilot Studio, they are used to present rich, interactive user interface elements directly within a conversation. This allows for the display of formatted text, images, input fields, and buttons, enhancing the user experience beyond simple text responses. They enable copilots to gather information from users or present complex data in a structured, visually appealing format without leaving the chat interface.

A. Adaptive Cards are for UI presentation, not permanent data storage. Data is typically stored in systems like Microsoft Dataverse or other external data sources.

C. APIs are managed through connectors and actions, not directly by Adaptive Cards, which are focused on the user interface layer of the conversation.

D. Message encryption is a function of the underlying chat platform (e.g., Microsoft Teams) and transport protocols, not a role of Adaptive Cards.

1. Microsoft Learn. (2024). Send a card in Copilot Studio. "You can use cards to present interactive messages to users, such as images, videos, buttons, and forms." See the "Prerequisites" and "Send a basic card" sections.

2. Microsoft, Adaptive Cards Documentation. (n.d.). Overview. "Adaptive Cards are platform-agnostic snippets of UI, authored in JSON, that apps and services can openly exchange." See the "What are Adaptive Cards?" section.

AI agents are designed to be more than just conversational bots; they can perform tasks in the real world. This is achieved by equipping them with "tools" or defining "actions" they can invoke. These tools are functions or APIs that the agent's underlying language model can choose to call based on the user's request. For example, an agent could have a tool to send an email, query a database, or book a meeting. This capability transforms the agent from a passive information provider into an active assistant that can execute tasks on behalf of the user, directly interacting with other systems and services.

A. Static responses are pre-defined answers and do not involve performing dynamic, real-world actions.

C. Logs are records of an agent's activity used for monitoring and debugging, not for performing actions.

D. Variables are used to store and manage state or data within the agent's logic, not to execute external tasks.

1. Microsoft Learn, "Add tools to an AI agent". This document explicitly states, "Tools are capabilities that you can enable for your agent, which allow it to perform a wider range of tasks." It details how to add tools like web search or custom Python functions. (Section: "Add tools to an AI agent")

2. Microsoft Learn, "Semantic Kernel - Add native functions to your agent". Semantic Kernel is a core Microsoft AI agent framework. This page explains how to create "plugins" with functions (actions) that the kernel can orchestrate. (Section: "Create a plugin with native functions")

In a multi-agent solution, where multiple specialized copilots or agents collaborate, success hinges on clear orchestration. Defining distinct roles for each agent (e.g., one for sales, one for support) and establishing precise interaction protocols (how they pass information or hand off conversations) is critical. This structured design prevents ambiguity, avoids conflicting actions, and ensures that user requests are routed to the correct agent efficiently. Without clear roles and interactions, the system would be chaotic and unable to perform complex tasks that require collaboration.

A. Ignoring communication is the primary cause of failure in any multi-agent system; clear communication protocols are essential for collaboration.

C. Using one agent only contradicts the fundamental premise of designing a "multi-agent" solution, which inherently involves more than one agent.

D. Disabling APIs would likely cripple a multi-agent system, as APIs are a common mechanism for agents to communicate and interact with each other.

1. Microsoft Learn. "Overview of generative AI in Microsoft Copilot Studio." Microsoft Copilot Studio Documentation. The section on "Copilot" describes how a single copilot can be extended with skills or plugins, which is a form of multi-agent design where each skill has a defined role.

2. Microsoft Learn. "Route conversations to a bot framework skill." Microsoft Copilot Studio Documentation. This document details a practical implementation of a multi-agent pattern, where a primary copilot (role: router) interacts with a specialized skill (role: specialist), demonstrating the need for defined roles and interactions.

The standard and most effective method for evaluating AI agent performance is to create a curated test set and systematically review the agent's results against it. This involves designing a collection of prompts and expected outcomes that cover a range of scenarios, including common user queries, edge cases, and potential failure points. By comparing the agent's actual responses to the expected ones, developers can quantitatively and qualitatively measure accuracy, relevance, and overall quality. This structured approach provides actionable feedback for iterative improvement and ensures the agent meets its performance goals before deployment.

A. Random testing lacks the systematic coverage of a curated test set and may miss critical failure modes or performance issues.

C. Ignoring outputs is the opposite of evaluation; it provides no information about the agent's performance or areas for improvement.

D. Manual guessing is an unstructured and unreliable approach that does not provide objective or repeatable metrics for performance evaluation.

1. Microsoft Learn. (2024). Test your Microsoft Copilot Studio copilot. Microsoft Copilot Studio Documentation. Section: "Test your copilot," Paragraph 1. "As you design your copilot, you can use the Test your copilot pane to see how the conversation flows." This pane allows for systematic review of results.

2. Google Developers. (2024). Testing and Evaluation. People + AI Guidebook. Chapter: "Testing and Evaluation," Section: "Create a test set." "Create a test set that represents the diversity of your intended users and use cases." This is a standard industry practice applicable to all AI agents.

Adaptive Cards are the standard feature for presenting structured forms and rich, interactive UI elements within an AI agent conversation. They are an open card exchange format that allows developers to define UI content in a common, declarative JSON format. The host application, such as Microsoft Teams or a custom website hosting a copilot, renders the card natively. This enables the agent to collect structured data from users—such as dates, choices, and text inputs—in a user-friendly and organized manner, far superior to a series of simple text prompts.

A. Variables are used to store and manage data within the conversation's logic, not to present user interface forms.

C. Logs are records of events and conversations used for debugging and analytics, not for user interaction.

D. Static prompts are simple text-based questions and lack the rich, interactive capabilities of a structured form.

1. Microsoft Copilot Studio Documentation. (n.d.). Send a card in Microsoft Copilot Studio. Microsoft Learn. Retrieved from https://learn.microsoft.com/en-us/microsoft-copilot-studio/advanced-cards. (This page explicitly details how to use Adaptive Cards to present information and collect input from users).

2. Adaptive Cards Documentation. (n.d.). Overview. adaptivecards.io. Retrieved from https://adaptivecards.io/docs/. (This is the official vendor documentation for the Adaptive Card framework, explaining its purpose for creating platform-agnostic UI snippets).

In Microsoft Copilot Studio, orchestrating a multi-step AI workflow involves a combination of conversational logic within topics and backend processes executed by flows. A topic structures the user interaction, and for complex actions, it calls a Power Automate flow. This flow can connect to hundreds of data sources, execute business logic, and return results to the copilot. The term "Agent flows" best describes this end-to-end orchestration, where conversational steps and automated actions are combined to fulfill a user's request.

A. "Topics only" is incomplete, as complex workflows often require Power Automate flows for actions.

C. Static responses are simple, single-turn outputs and cannot orchestrate a multi-step process.

D. Logs are records of execution used for auditing and debugging, not for orchestrating workflows.

1. Microsoft Learn. (2024). Use Power Automate flows in Microsoft Copilot Studio. "You can enable your copilot to perform actions by calling a Power Automate flow. Flows can help you automate activities, or call backend systems." Section: "Create a new flow from the authoring canvas".

2. Microsoft Learn. (2024). Create and edit topics in Microsoft Copilot Studio. "A topic defines a conversation path... For complex scenarios, you can call Power Automate flows." Section: "Create a topic".

In AI agent solutions, particularly those using the Retrieval-Augmented Generation (RAG) pattern, Azure AI Search serves as the specialized data store. Its role is to ingest, index, and enrich an organization's private data (e.g., documents, manuals, databases). The AI agent can then query this indexed data to retrieve relevant, factual information to ground its responses. This allows the agent to answer questions based on specific, up-to-date, and proprietary information that was not part of the foundational model's original training data, enhancing accuracy and relevance.

A. UI rendering: Azure AI Search is a backend data service; it does not handle front-end user interface rendering.

C. Managing authentication: Authentication is primarily managed by Microsoft Entra ID, although AI Search integrates with it to enforce access controls on the data.

D. Hosting APIs: While it provides a query API, its purpose is specific to search and retrieval, not general-purpose API hosting like Azure API Management.

1. Microsoft Azure AI Documentation. (2024). Retrieval-Augmented Generation (RAG) in Azure AI Search. Microsoft Learn. This document explicitly states, "Azure AI Search is a central component in RAG applications," detailing its role in data ingestion, indexing, and information retrieval to support LLMs. See the "RAG architecture and workflow" section.

2. Microsoft. (2024). Azure AI Search Overview. Microsoft Learn. The documentation's introductory section describes the service as a "fully managed search-as-a-service" for adding rich search experiences over private, heterogeneous content, which is its exact function in a RAG-based agent.

The combination of Azure AI Search and Retrieval-Augmented Generation (RAG) is a powerful feature that enables AI agents to process vast amounts of both structured (e.g., databases) and unstructured (e.g., PDFs, websites, Word documents) data. Azure AI Search indexes the data, making it searchable. The RAG pattern then allows the AI agent to retrieve relevant information from this index in real-time to generate accurate, context-aware responses, effectively grounding the agent in a specific knowledge base.

Static topics are pre-scripted conversation paths and are not designed to dynamically process large, varied data sources.

Manual scripts are a generic solution; Azure AI Search with RAG is a specific, integrated, and scalable Microsoft feature for this purpose.

Logs are output records of system activity used for diagnostics and analytics, not for processing input data to answer user queries.

1. Microsoft Learn. (2024). Use a search and summarize content action. "You can use the Azure AI Search connector to have your copilot search for information in an Azure AI Search index... This feature uses a technique called Retrieval-Augmented Generation (RAG)."

2. Microsoft Learn. (2024). Generative answers with Search and summarize. "This capability lets your copilot use information from multiple sources... to answer user questions without you having to create topics... It can use both unstructured documents and structured data sources." Section: Data sources.

3. Azure Architecture Center. (2024). Retrieval-Augmented Generation (RAG) in Azure AI Search. "RAG patterns supplement prompts with external data, which is retrieved from a data source such as a search index... Azure AI Search can provide the data retrieval system in a RAG pattern."

Application Lifecycle Management (ALM) is the process that ensures AI agents and their components are moved between environments (e.g., development, testing, production) in a consistent and reliable manner. In the context of Microsoft Copilot Studio, ALM is implemented by packaging copilots, topics, flows, and other dependencies into "solutions." These solutions can then be exported from a source environment and imported into a target environment. This automated and standardized process eliminates manual configuration errors and guarantees that the agent behaves identically across all stages of its lifecycle.

A. Manual configuration is error-prone and the primary cause of inconsistencies between environments.

C. Static prompts are only one small part of a copilot's configuration; ALM manages the entire asset.

D. Logs are used for monitoring and diagnostics after deployment, not for ensuring deployment consistency.

1. Microsoft Learn. (2024). Implement ALM for Microsoft Copilot Studio copilots. "Application lifecycle management (ALM) is important for any business... You can use solutions to export and import your copilots... to move your copilots across multiple environments." Section: "Solutions".

2. Microsoft Power Platform documentation. (2024). Application lifecycle management (ALM) with Microsoft Power Platform. "ALM is the lifecycle management of applications, which includes governance, development, and maintenance... Power Platform provides a set of tools... to support ALM." Section: "Foundations of ALM".