AB-620: Microsoft AI Agent Builder Associate Exam Prep

Lists the accounts and licences the lab needs, and why personal email fails.

Create a tenant you administer and the work account you will build with.

Sets out which environment your lab work belongs in and why it matters.

Start the Copilot Studio trial at the right moment and clear sign-up blocks.

Prepare Azure for later labs and set a budget alert that warns of any spend.

Import the Fernwood Outfitters records and documents the labs build on.

Spot which Copilot Studio experience you are in and why labs use standard.

A five-way rule for choosing instructions, knowledge, tools, topics or triggers.

Know when a scenario calls for deterministic topics or model-led planning.

Split declarative from custom engine and see where Copilot Studio agents fall.

The Power Fx you need: scope prefixes, decimal separators and a worked formula.

Add Dataverse tables as knowledge, enable Dataverse search and tune answers.

Where things live, what travels and whose they are: the basis of ALM.

See how cards display and collect structured data, and what breaks by host.

See how domain one splits between planning and hands-on building.

Four questions to answer before connecting an agent to any business system.

A decision rule for picking the most governable integration pattern.

Decide between the three authentication options and when each is required.

Configure each sign-in option and keep user auth apart from agent identity.

Choose a channel by working back from who the users are and how they sign in.

Two traps: Teams SSO outside one-to-one chats and SharePoint without sign-in.

Grounding, moderation and citations, and when to stop answers beyond sources.

A four-question security review covering access, reach, injection and audit.

Use environments, Managed Environments and data policies to enforce rules.

Where connector policies end and Purview's label-based protection begins.

Find every agent in the tenant, read its metadata and block one if needed.

A rule for what to share across agents and what reuse really costs.

Test whether the agent knows its user, then set scope and safety to suit.

Tell the three apart by where they were authored, not by what they are called.

Build a flow that replies within about 100 seconds, with async response off.

Design the flow's inputs and outputs as a clear contract with the agent.

Choose action types well and see why premium connectors can fail at run time.

Pause a flow for a human decision without leaving the agent waiting.

Know the three documented failure shapes and handle idempotency yourself.

Trace a failed run step by step and tell flow runs apart from agent telemetry.

Choose between attaching a flow as a tool and calling it from a topic.

Build a topic when the sequence matters and let orchestration handle the rest.

Write distinct trigger phrases, configure question nodes and cover every branch.

Pick the right variable scope and know why ending all topics leaves globals set.

Call a flow at a chosen point in a topic and diagnose why it is not listed.

Attach connector actions, prompts and MCP tools through one topic entry point.

Set node-level sources, which outrank the agent's, and keep citations intact.

Format messages lightly and plan for rendering differences between channels.

Build a prompt with a defined output shape, ground it in Dataverse, and call it.

Feed runtime data to generative answers and see how much of it is used.

Call any web endpoint from a topic, parse the response and handle failures.

Build cards to the lowest schema version your target channels support.

Collect structured input with a card and turn submissions into variables.

Plan the Fernwood Outfitters returns agent and build its first working topic.

Map the knowledge, tools, multi-agent and Azure work ahead in the second domain.

Match each need to a source: records or documents, indexed or live.

Wire the three most used sources and know which content belongs in each.

Tell connectors that let an agent know things from those that let it act.

See which setup steps an administrator owns and the short part that is yours.

Query a system live under the user's identity, with no content copied.

Learn indexes, chunking and hybrid search, and why retrieval decides quality.

Connect a vector index to an agent safely and know which billing meters apply.

Explain why agents expose oversharing and fix permissions at the source.

Choose among five kinds of tool, preferring the most governable that fits.

Add a connector tool, choose maker or user credentials and fix its description.

Turn an API specification into a governed connector in the Power Platform tools.

Add only the actions you need and test that the orchestrator selects them.

Place a REST API on the spectrum from custom connector to a single HTTP call.

Decide whose identity a tool uses and keep secrets out of exported solutions.

Describe the situation a tool serves so the orchestrator can tell tools apart.

Tell a fixed connector apart from an MCP server whose tools are discovered live.

Add an MCP server, choose its authentication and check why a connection fails.

Work with server-owned tool descriptions and switch on only the tools you need.

Know who allows a server for the tenant and who picks the tools an agent uses.

Tell computer use apart from recorded RPA, and know where it can run.

Add a computer use tool, write outcome-based instructions and set an allow-list.

Rule out poor fits for computer use, and know the limits of human review.

Read step-by-step run records, logs and credit use to see how runs behave.