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

STEP 1 · Study

AB-620 StudyAI Agent Builder (Copilot Studio)

Read the key points for each exam domain, then use “Practice this domain” to answer questions from that domain only. Domains with a larger weight appear more often on the exam.

01

Plan and configure agent solutions

30–35%

Learning goals

  • Plan authentication, channels and publishing scope separately for internal and external agents
  • Design agent flows that include human-in-the-loop approval and error handling
  • Choose between topics, variables, Adaptive Cards and generative answers nodes by purpose

Key points

  • Decide first who will use the agent. Internal (employees) or external (customers) determines the authentication method, publishing channels and which knowledge it may use.
  • Identity strategy: state explicitly whether documents are accessed with the user’s own permissions or through the agent’s connection. A design that mixes documents the user cannot view into answers is not acceptable.
  • Agent flows define input and output parameters and exchange them with topic variables. Put an approval step (human-in-the-loop) in front of high-value or high-impact actions.
  • Error handling: the basic rule is “if it fails, stop and notify a person.” Decide the retry count and timeout, and never report a failure as success.
  • Mind variable scope: use topic variables within one topic, global variables to share across topics, and environment variables for values that differ by environment.
  • Separate roles: topics for fixed branching, generative answers nodes for free-form questions grounded in policy documents, and Adaptive Cards to return tables or buttons.
  • Governance: decide on separate dev/test/production environments, connector control with DLP policies and shared reusable components at the planning stage.

Workplace uses

  • SalesQuote inquiry topic: take the customer name, retrieve the quote status from the core system with an agent flow and show it in an Adaptive Card. Discount changes go to a manager-approval flow.
  • General affairsOffer an internal help desk in Teams. Answer questions about work rules and equipment requests with a generative answers node grounded in policies, and take requests in a fixed topic.
  • FinancePayment request flow: require department-head approval with an approval action above a set amount. On failure, stop and notify the finance owner.

Exam objectives (from the exam guide)

  • Plan an agent solution
  • Create and monitor agent flows in Copilot Studio
  • Configure topics

Learning resources

Learn in detail (3 topics with mini quizzes) →Practice this domain (100 questions) →Next domain ↓
02

Integrate and extend agents in Copilot Studio

40–45%

Learning goals

  • Explain when to use Copilot connectors, Power Platform connectors and Azure AI Search
  • Decide when to add MCP tools, custom connectors, REST APIs and computer use as tools
  • Design multi-agent solutions with Foundry agents, Fabric data agents and A2A
  • Understand monitoring with Application Insights and use of the Foundry model catalog

Key points

  • Separate knowledge (reference) from tools (actions). Searching and grounding on documents in external systems is knowledge; operations such as creating or updating records are tools.
  • Copilot connectors bring external content into Microsoft 365 for reference; Power Platform connectors are mainly for data operations and actions in external services.
  • For large document sets that already have an Azure AI Search index, connecting Azure AI Search as knowledge is the shortest path.
  • MCP (Model Context Protocol) is a standard for providing tools and data to agents. Tools are defined on the server and can be reused by multiple agents.
  • A2A (Agent2Agent) is a protocol for agents to collaborate with each other. Do not confuse it with MCP (tool connection).
  • Legacy systems without APIs can be automated through the UI with computer use, but this assumes a limited scope, monitoring of runs and human review.
  • For REST APIs and custom connectors, the clearer the operation descriptions and inputs/outputs, the more accurately orchestration chooses them. Never hard-code credentials in prompts or code.
  • Send production conversation and error telemetry to Application Insights to fit into existing monitoring operations.

Workplace uses

  • SalesUse a data agent to aggregate “orders by region this quarter” from sales data in Fabric, and call it from the sales agent.
  • General affairsBring ServiceNow knowledge articles in with a Copilot connector and use them for first-line answers to internal inquiries.
  • FinanceTrial computer use for entering journal entries into old accounting software without an API. A staff member approves on a confirmation screen before entry.

Exam objectives (from the exam guide)

  • Connect to enterprise knowledge sources
  • Add tools to agents
  • Configure multi-agent collaboration from Copilot Studio
  • Integrate agents with Azure

Learning resources

Learn in detail (4 topics with mini quizzes) →Practice this domain (131 questions) →Next domain ↓
03

Test and manage agents

20–25%

Learning goals

  • Create test sets and choose evaluation methods that fit the nature of the questions
  • Classify failure causes from test results and turn them into improvements
  • Build ALM with solutions, environment variables and Power Platform Pipelines

Key points

  • A test set pairs representative questions with expected answers (or conditions the answer must meet). Rerun it after every change to check for regressions.
  • Choose the evaluation method by the nature of the question: exact match when there is one correct answer, and semantic similarity or answer-quality evaluation for free-form answers.
  • Classify failures by cause (missing knowledge, wrong topic routing, tool call failure, insufficient permissions, etc.). Each cause needs a different fix.
  • Put agents in solutions to move them between environments. Existing agents can be added to a solution later.
  • Externalize values that differ by environment, such as API URLs, into environment variables, and do not edit production directly.
  • Automate dev → test → production deployment with Power Platform Pipelines, and extend it with approvals as needed.

Workplace uses

  • SalesRun the same 30-question test set before and after changing the quote inquiry agent, and compare accuracy and the types of wrong answers.
  • General affairsWhen policies are revised, check for regressions with a test set that adds only the related questions, then deploy to production through the pipeline.
  • FinanceSwitch the accounting API endpoint between dev and production with environment variables to prevent accidental connections to production data.

Exam objectives (from the exam guide)

  • Evaluate agent performance
  • Implement application lifecycle management (ALM) for agents in Copilot Studio

Learning resources

Learn in detail (2 topics with mini quizzes) →Practice this domain (70 questions) →Try the mock exam →

When you finish reading

Practice all domains (301 questions)Mock exam (50 questions · 75 min)AB-620 home
Notes (unofficial, original questions, no pass guarantee)