AB-620 · CHEAT SHEET

Microsoft Certified: AI Agent Builder Associate

Key concepts grouped by official exam weight.

1. Plan and configure agent solutions

30–35%

Design how the agent connects to enterprise systems, selecting connectors, APIs, flows, or other patterns based on security, latency, and maintainability.
Define authentication, authorization, and user or author identity so every interaction applies appropriate permissions and least privilege.
Choose channels and deployment environments based on audience, authentication, supported capabilities, governance, and publishing lifecycle.
Plan responsible AI controls for safety, transparency, privacy, human oversight, usage boundaries, and continuous agent evaluation.
Assess security and governance risks, including data policies, permissions, information exposure, connectors, and channel publishing.

Create an agent flow with well-defined triggers, inputs, logic, actions, and outputs to automate deterministic processes.
Add human approval or intervention at high-impact decision points, exceptions, or actions that require review.
Configure actions and connectors with clear contracts, valid connections, appropriate authentication, and parameters the agent can use reliably.
Monitor flow runs, duration, errors, and dependencies to detect failures, bottlenecks, and operational regressions.
Define input and output parameters with unambiguous names, types, and descriptions to exchange data correctly between agent and flow.

Add an agent flow to a topic, map its inputs and outputs, and use the results in later conversation steps.
Format agent responses using text, lists, tables, variables, or structures appropriate to the channel and user need.
Add tools within a topic and configure when they run, what data they receive, and how their results are processed.
Create custom prompts with appropriate instructions, inputs, outputs, model, and testing to produce advanced, consistent responses.
Integrate custom knowledge sources and transform results into the expected format to ground controlled generative answers.

40–45%

Connect indexed external enterprise sources through Microsoft Copilot connectors while preserving permissions and configuring required authentication.
Use Power Platform connectors as tools or real-time knowledge by choosing appropriate connections, credentials, tables, and policies.
Connect Azure AI Search to retrieve relevant enterprise index content and ground agent responses with authorized results.

Configure computer use to automate web or Windows applications without a suitable API and monitor its runs, credentials, and failures.
Add MCP servers as tools, authorize connections, review resources, and limit enabled tools according to need and governance.
Reuse existing custom connectors as tools by configuring actions, parameters, authentication, connection references, and dependencies.
Add a REST API through a clear contract, well-described operations, stable schemas, secure authentication, and structured responses.

Design multi-agent solutions with separated responsibilities, clear delegation criteria, least privilege, observability, and controlled context handoff.
Connect a Microsoft Foundry agent through its project and agent ID, define routing metadata, and validate security, quality, and data handling.
Connect an existing Copilot Studio agent that is published, accessible, and in the same environment, configuring its description and shared history.
Integrate a published Fabric data agent to answer over OneLake data, choosing user or maker authentication according to permission requirements.
Use the Agent2Agent protocol to connect and orchestrate compatible external agents, configure authentication, and validate end-to-end delegation.

Combine Azure AI Search with Microsoft Foundry to retrieve relevant knowledge and generate grounded responses through an evaluable, secure architecture.
Configure prompts with Microsoft Foundry catalog models using supported deployments, chat completions endpoints, and compatible modalities.
Integrate Application Insights to correlate requests, dependencies, duration, exceptions, and failures and diagnose agent behavior.

20–25%

Create representative test sets with prompts, expected answers, acceptance criteria, profiles, and positive, negative, and risk cases.
Choose evaluation methods such as text match, similarity, or general quality depending on whether the expected outcome is exact, semantic, or open-ended.
Analyze passes, failures, errors, responses, activity, latency, and run comparisons to detect causes and regressions.

Create an unmanaged development solution with an appropriate publisher and version to group, maintain, and transport the agent and its components.
Add existing agents and all required objects to solutions, checking topics, flows, connectors, references, and dependencies before export.
Use environment variables to separate environment-specific values from solution design by managing definitions, default values, and current values.
Implement and extend Power Platform Pipelines for ordered deployments, preflight validation, approvals, target values, and centralized governance.