CertSafari

    Free Microsoft Power Platform AI Agent Builder Associate (AB-620) - December 18 onwards Sample Questions

    35 free sample questions from our bank of 352+, covering every exam domain, with answers and detailed explanations. Updated November 2026.

    1: Plan and configure agent solutions

    1.3: Create and manage skills

    1.When a skill fails a load-time check, Copilot Studio ______ the affected skill and still loads the rest of the agent.

    1. A.skips
    2. B.deletes
    3. C.republishes
    Show answer & explanation

    Correct answer: A — skips

    • A. Correct: the failing skill is skipped while the remainder of the agent loads.
    • B. Incorrect: nothing is deleted; the file simply isn't loaded.
    • C. Incorrect: republishing is unrelated to load-time validation.

    1.1: Plan an agent solution

    2.A planner reviews this tool configuration for an agent that reads salary bands from an HR system for managers only: ``` name: HRSalaryBands transport: streamable-http url: https://hr.contoso.com/mcp authentication: none run_as: agent_author ``` Which setting is at fault for the stated requirement?

    1. A.name: HRSalaryBands
    2. B.transport: streamable-http
    3. C.run_as: agent_author
    4. D.url: https://hr.contoso.com/mcp
    Show answer & explanation

    Correct answer: C — run_as: agent_author

    • A. The name is only a label; changing it does nothing to enforce who can see salary data.
    • B. Streamable HTTP is a supported MCP transport and is not the cause of the access problem.
    • C. Running as the author means every user gets the maker's access to salary data; the tool should run with each user's credentials so HR permissions apply.
    • D. An HTTPS endpoint on the company domain is acceptable; the URL is not the flaw.

    1.4: Configure agent identity, instructions, and behavior

    3.A maker built an agent on the Microsoft Copilot harness with Memory turned on. During a compliance review the maker switches the Memory toggle off on the Build tab. What happens to the memories already stored for users?

    1. A.They are erased from every user's memory portal at the moment the toggle is switched to the off position.
    2. B.They keep being applied to answers, because only a tenant admin can stop an agent from reading saved notes.
    3. C.They are copied into the agent's knowledge sources so that makers can review them from the Knowledge page.
    4. D.They stay stored until deleted or until 28 days of inactivity pass, but the agent stops saving and using them.
    Show answer & explanation

    Correct answer: D — They stay stored until deleted or until 28 days of inactivity pass, but the agent stops saving and using them.

    • A. Incorrect: switching the toggle off is not a purge. Stored notes remain until users delete them or the inactivity period expires.
    • B. Incorrect: the maker controls whether the agent uses Memory through the toggle; no tenant admin action is needed to stop recall.
    • C. Incorrect: memories are never converted into knowledge sources, and personal memories are not visible to makers or other users.
    • D. Correct: turning Memory off stops the agent from saving or using memories, but it does not delete them right away; they age out after 28 days of inactivity or are removed in the portal.

    1.4: Configure agent identity, instructions, and behavior

    4.To keep a standard-harness agent on the latest supported GPT chat model as snapshots update, select the ___ alias as its primary model.

    1. A.GPT-4.1
    2. B.GPT Chat Latest
    3. C.GPT-5 Auto
    Show answer & explanation

    Correct answer: B — GPT Chat Latest

    • A. GPT-4.1 is a fixed model with a retirement date, not an alias that follows updates.
    • B. GPT Chat Latest is the alias that follows the latest supported chat model and is moved automatically when models retire.
    • C. GPT-5 Auto is a fixed model that routes dynamically and is retiring, not the alias for the latest chat model.

    1.3: Create and manage skills

    5.A maker edited the instructions of an uploaded skill in a text editor and wants the agent to use the changes. What is the documented way to apply them?

    1. A.Open the skill's three-dot menu, choose Replace, then browse to the edited file and select it
    2. B.Open the skill list, add the edited file as a new skill, then keep both skills enabled
    3. C.Publish the agent again from the Publish page, then start a new test conversation in Preview
    4. D.Save the edits over the downloaded file on disk, then wait for the agent to pick up the change
    Show answer & explanation

    Correct answer: A — Open the skill's three-dot menu, choose Replace, then browse to the edited file and select it

    • A. The Replace command in a skill's three-dot menu swaps the stored skill for the edited file, so the agent uses the new instructions in place of the old ones.
    • B. Adding the edited file creates a second, separate skill instead of updating the first. The original keeps its old instructions and stays enabled next to the new one.
    • C. Publishing only releases what is already stored in Copilot Studio. The stored skill still holds the old text because the edited file was never supplied to it, so a new test conversation behaves the same.
    • D. The uploaded skill is a copy held in Copilot Studio and has no link to the file on disk. Changing the local file leaves the agent untouched until the file is supplied again.

    1.3: Create and manage skills

    6.A maker deletes a skill from an agent on the Microsoft Copilot harness, then realizes it may be needed in a later project. What guidance applies?

    1. A.Deleting a skill removes it from that agent for good; download a copy beforehand to use it again
    2. B.Deleting a skill moves it to a recycle area for thirty days; restore it there to use it again
    3. C.Deleting a skill only hides it in the components panel; toggle it back on there to use it again
    4. D.Deleting a skill also removes the copies shared with others; rebuild it from scratch to use it again
    Show answer & explanation

    Correct answer: A — Deleting a skill removes it from that agent for good; download a copy beforehand to use it again

    • A. Removing a skill from an agent is permanent, with no undo or recovery step. The documented advice is to download the skill before deleting it if it might be wanted in a later project.
    • B. Skills have no recycle area and no thirty-day retention window. Once a skill is deleted from an agent there is nothing to restore, so a copy has to be saved beforehand.
    • C. Deleting a skill removes it rather than hiding it. Skills are deleted through a confirmation step, and there is no disabled state in the components panel that can be toggled back on.
    • D. Deleting a skill affects only the agent it is deleted from. Copies that other people added to their own agents are separate and stay in place, so there is no loss of shared copies to work around.

    1.3: Create and manage skills

    7.Case study: Fabrikam IT built a skill, then needs it reused by agents across the organization. Only the IT-ops group may use it and only the author may change it. Which action meets the requirement in Copilot Studio?

    1. A.Share the skill with the IT-ops group, whose members add it to their agents via Browse skills (preview)
    2. B.Export the skill as a ZIP file and send it to the IT-ops group, whose members upload it to their agents
    3. C.Publish the author's agent to Teams for the IT-ops group, whose members call the skill via that agent
    4. D.Make the author's agent editable by the IT-ops group, whose members reuse the skill from that agent
    Show answer & explanation

    Correct answer: A — Share the skill with the IT-ops group, whose members add it to their agents via Browse skills (preview)

    • A. Sharing the skill with the IT-ops group limits use to that group. Its members browse the skill and add it to their own agents, while edit rights remain with the author.
    • B. An exported ZIP uploaded by each recipient becomes an independent copy that the recipient owns and can modify. The author no longer controls changes, and later updates do not reach those copies.
    • C. Publishing the author's agent to Teams only lets people chat with that one agent. It does not let IT-ops add the skill to agents they build, so the skill is not reused across their agents.
    • D. Making the author's agent editable by IT-ops lets its members change the skill and the agent. That breaks the rule that only the author may change it, and the skill still does not reach their own agents.

    1.1: Plan an agent solution

    8.A team is planning environments for a customer-facing agent that uses a connection to Dynamics 365 and an API base URL that differs per stage. Which TWO planning decisions support promoting the agent from development to production?(Select 2)

    1. A.Recreate the agent in production by copying its topic text into a new agent, then add the Dynamics 365 connection by hand there
    2. B.Import the solution as managed, then edit the topics directly in production to set the production URL and Dynamics 365 connection
    3. C.Create the agent in a custom solution rather than the default solution, then export that solution from development for import elsewhere
    4. D.Keep the API base URL in a global variable inside the agent and reuse one Dynamics 365 connection in development and production
    5. E.Keep the API base URL in an environment variable and the Dynamics 365 link in a connection reference, set separately per environment
    Show answer & explanation

    Correct answers: C, E — Create the agent in a custom solution rather than the default solution, then export that solution from development for import elsewhere; Keep the API base URL in an environment variable and the Dynamics 365 link in a connection reference, set separately per environment

    • A. Rebuilding the agent by hand is not promotion. It leaves production as a separate copy that drifts from development and has to be repeated at every release. Solutions can carry the connection reference, so the manual connection step is unnecessary.
    • B. Components in a managed solution are meant to stay unchanged in the target. Editing them in production adds unmanaged customizations that cause conflicts on later imports. Per-stage values belong in environment variables and connection references.
    • C. Agents have to live in a custom solution to be exported and imported. The default solution is not a transportable package, so the custom solution is the vehicle for moving the agent from development to production.
    • D. A global variable lives inside the agent and travels with it, so production would receive the development URL and need an edit after import. Reusing one connection also risks pointing development tests at production data. The URL differs per stage, so it needs an environment variable instead.
    • E. An environment variable holds the stage-specific API base URL, and a connection reference points to the Dynamics 365 connection. The same solution can then be imported into each environment and bound to that environment's own values.

    1.1: Plan an agent solution

    9.The security team wants a central record of maker activity on agents and automated alerting when suspicious agent activity occurs. Which combination should be planned?

    1. A.Conversation transcripts in Dataverse for the record and Application Insights alert rules for suspicious agent activity
    2. B.Customer Lockbox request history for the record and Copilot credit consumption alerts for suspicious agent activity
    3. C.Maker audit logs in Microsoft Purview for the record and Microsoft Sentinel for alerts on suspicious agent activity
    4. D.Power Platform solution checker results for the record and Teams admin center alerts for suspicious agent activity
    Show answer & explanation

    Correct answer: C — Maker audit logs in Microsoft Purview for the record and Microsoft Sentinel for alerts on suspicious agent activity

    • A. Dataverse transcripts capture conversations between users and the agent, not the changes makers make to it, so they cannot serve as a record of maker activity. Application Insights alert rules work on runtime telemetry, not on security events.
    • B. Customer Lockbox history covers requests for Microsoft engineers to access tenant data, not what makers do to agents. Credit consumption alerts track spending and usage, so they do not flag suspicious behavior.
    • C. Purview holds the audit logs of maker actions on agents, which gives the security team a central record. Sentinel can ingest that activity, monitor it and raise automated alerts on suspicious agent activity.
    • D. Solution checker results describe the quality and compliance of solution components, not who changed an agent. The Teams admin center manages Teams settings and does not provide security alerting on agent activity.

    1.1: Plan an agent solution

    10.Which authentication option is turned on by default for a new Copilot Studio agent?

    1. A.Authenticate manually with OAuth 2.0
    2. B.Authenticate manually with Microsoft Entra ID
    3. C.No authentication
    4. D.Authenticate with Microsoft
    Show answer & explanation

    Correct answer: D — Authenticate with Microsoft

    • A. Manual authentication with a generic OAuth 2.0 provider is a real Copilot Studio setting, but the maker has to select and configure it. A new agent does not start with it.
    • B. Manual authentication can use Microsoft Entra ID as its provider, but the maker must choose it and register the app. A new agent starts on the built-in Microsoft option, not on manual configuration.
    • C. Anonymous access must be chosen deliberately.
    • D. It is the default and uses Entra ID for Teams, Power Apps and Microsoft Copilot.

    1.4: Configure agent identity, instructions, and behavior

    11.A cost-sensitive FAQ agent drafts short answers, summarizes documents and translates replies from a grounded knowledge source. Which TWO reasons support selecting a General-tagged model? (Choose 2.)(Select 2)

    1. A.General models are retired on a fixed quarterly cycle and replaced by an Auto model.
    2. B.General models plan and run multistep work across many tools, with deliberate reasoning.
    3. C.General models pick a model size for each query, balancing answer quality against cost.
    4. D.General models offer the lowest latency and cost, with shallow-to-moderate reasoning depth.
    5. E.General models are tuned for speed and cost on everyday chat and light grounding.
    Show answer & explanation

    Correct answers: D, E — General models offer the lowest latency and cost, with shallow-to-moderate reasoning depth.; General models are tuned for speed and cost on everyday chat and light grounding.

    • A. Incorrect: General models have no fixed quarterly retirement cycle with an Auto replacement. Each model is retired on its own published date, and lifecycle timing says nothing about fit for short drafting, summarizing and translation.
    • B. Incorrect: planning and running multistep work across many tools with deliberate reasoning is the Deep profile. Short answers from a single grounded knowledge source do not call for it.
    • C. Incorrect: choosing a model size per query is what Auto models do. General is a fixed tier with a stable speed and cost profile, not a router, so it does not give that benefit.
    • D. Correct: General offers the lowest latency and cost and accepts shallow-to-moderate reasoning depth. That trade-off suits a cost-sensitive FAQ agent whose tasks are short and simple.
    • E. Correct: General is tuned for speed and cost on everyday chat and light grounding. That matches drafting, summarizing and translating replies from a grounded knowledge source.

    1.4: Configure agent identity, instructions, and behavior

    12.A telecom plans an agent that remembers returning customers' preferences, uses reusable skills, and is published on a phone number with escalation to human agents. Which TWO statements about harnesses are correct? (Choose 2.)(Select 2)

    1. A.The Customer Conversation harness offers skills and memory, and agents built on it can be published to a phone line or web chat.
    2. B.An agent on the Microsoft Copilot harness can be moved back to the standard harness, and the switch is a single toggle.
    3. C.The Microsoft Copilot harness offers skills and memory, and skills take over the role of the standard harness's prompt tool.
    4. D.The standard harness offers memory, and it is enabled by turning on generative orchestration in Generative AI settings.
    5. E.The Microsoft Copilot harness offers manual or custom user authentication, and it is configured as in the standard harness.
    Show answer & explanation

    Correct answers: A, C — The Customer Conversation harness offers skills and memory, and agents built on it can be published to a phone line or web chat.; The Microsoft Copilot harness offers skills and memory, and skills take over the role of the standard harness's prompt tool.

    • A. Correct: the Customer Conversation harness targets voice and chat customer service, supports both skills and memory, and agents on it can be published to a phone number or website chat with escalation to human agents.
    • B. Incorrect: the move is one-way. An agent created on the Microsoft Copilot harness can't be transferred back to the standard harness, and no setting reverses it.
    • C. Correct: the Microsoft Copilot harness supports skills and memory, and in it skills take the place of the prompt tool that the standard harness uses.
    • D. Incorrect: the standard harness doesn't support memory. Turning on generative orchestration in Generative AI settings doesn't change that, so it can't deliver the remembered customer preferences.
    • E. Incorrect: the Microsoft Copilot harness doesn't support manual or custom user authentication. The standard harness does, so the two harnesses are not configured alike on this point.

    1.2: Create and configure workflows

    13.A scheduled agent flow that syncs invoices failed overnight. The maker wants to see which action failed and what it received and returned. Where should they look first?

    1. A.Open Flow checker in the designer, select the warning on the failed step, and read the action's inputs and outputs
    2. B.Open run history on the flow's details page, select the failed run, and read the red action's inputs and outputs
    3. C.Open Version history on the flow's details page, select the newest version, and read the failed action's inputs and outputs
    4. D.Open the agent's test chat, turn on the Activity map, and select the failing tool node to read its inputs and outputs
    Show answer & explanation

    Correct answer: B — Open run history on the flow's details page, select the failed run, and read the red action's inputs and outputs

    • A. Flow checker validates the flow's design and reports configuration issues before saving or publishing. It holds no data from past executions, so it cannot show what an action received or returned during the overnight failure.
    • B. Run history on the flow's details page lists past runs with their status. Opening the failed run shows each action, and expanding the failed one reveals its inputs, outputs and error details.
    • C. Version history lists saved definitions of the flow so an earlier design can be compared or restored. It does not store the results of individual runs, so it shows no action inputs or outputs from the failed execution.
    • D. The Activity map visualises what happens during a conversation in the test chat. A scheduled run that no conversation triggered never appears there, so the failing action cannot be inspected from it.

    1.2: Create and configure workflows

    14.A maker wants a nightly-built agent flow to be usable as an agent-level tool. Select the THREE requirements it must meet.(Select 3)

    1. A.The flow starts from a Recurrence trigger, and the agent reads the result of its latest scheduled run
    2. B.The flow starts from the When an agent calls the flow trigger and contains a Respond to the agent action
    3. C.The flow is saved in the environment's Default Solution, not in a custom solution of that environment
    4. D.The flow includes at least one premium connector action alongside its trigger and response steps
    5. E.The flow has been published in its environment after its last edit, rather than only saved as a draft
    6. F.The flow's Respond to the agent action has Asynchronous response switched off in its Networking settings
    Show answer & explanation

    Correct answers: B, E, F — The flow starts from the When an agent calls the flow trigger and contains a Respond to the agent action; The flow has been published in its environment after its last edit, rather than only saved as a draft; The flow's Respond to the agent action has Asynchronous response switched off in its Networking settings

    • A. A Recurrence trigger only starts runs on a schedule. A flow used as an agent tool must start from the agent trigger and reply to the agent, so a scheduled flow is not offered as a tool.
    • B. The agent trigger and the Respond to the agent action define the flow's contract with the agent. A flow without both is not listed when adding a tool.
    • C. Flows in custom solutions, which are the recommended container, can be added as tools just as well. Living in the Default Solution is not a requirement for being listed.
    • D. Connector tier plays no part in tool eligibility. A flow built only from standard or built-in actions is as valid an agent tool as one with premium actions.
    • E. Only a published flow can be attached to the agent as a tool. A flow that exists only as a draft with unpublished edits does not appear for the agent.
    • F. The agent waits for the flow's reply, so the response must be synchronous. Leaving Asynchronous response on under Networking makes the flow unsuitable as a tool.

    1.2: Create and configure workflows

    15.Which TWO statements are documented limitations of multistage approvals in agent flows?(Select 2)

    1. A.Approval requests from multistage flows do not appear in Teams or Outlook, so approvers must respond in the Power Automate portal
    2. B.Conditions are not supported between stages, so routing logic has to sit in separate Condition actions outside the approval
    3. C.An approver cannot be assigned to more than one stage of a multistage approval, so the flow fails if one person is in two stages
    4. D.Advanced approvals do not support ALM, so a flow moved to another environment needs its multistage approval rebuilt
    5. E.A multistage approval supports one manual stage and one AI stage, so a third stage cannot be added to the flow
    Show answer & explanation

    Correct answers: C, D — An approver cannot be assigned to more than one stage of a multistage approval, so the flow fails if one person is in two stages; Advanced approvals do not support ALM, so a flow moved to another environment needs its multistage approval rebuilt

    • A. Multistage requests reach approvers in the Teams approvals app and Outlook as well as the Power Automate portal, as with standard approvals. This is not a documented limitation.
    • B. Conditions between stages are a core capability. They decide whether the approval ends as approved or rejected, or moves on to another stage, so no outside Condition actions are needed. This is not a limitation.
    • C. Assigning the same approver to more than one stage of a single multistage approval is not supported and makes the flow fail. The documentation lists this in the stage configuration note and in the known limitations.
    • D. The documentation lists the lack of ALM support for advanced approvals as a known limitation. Moving a flow between environments does not carry the approval definition, so it has to be rebuilt.
    • E. The documentation shows examples that combine several manual and AI stages with conditions. No cap of one stage of each type exists, so this is not a documented limitation.

    1.2: Create and configure workflows

    16.A maker wants to customise the retry behaviour for an HTTP action that calls a rate-limited vendor API. Select the TWO retry policy types that the action Settings provide.(Select 2)

    1. A.Linear interval, which raises the wait between consecutive attempts by a constant step
    2. B.Exponential interval, which multiplies the wait between consecutive attempts after a failure
    3. C.Fixed interval, which keeps the wait between consecutive attempts at one constant length
    4. D.Jittered interval, which randomises the wait between consecutive attempts within a set window
    5. E.Circuit breaker, which pauses calls to the vendor after a set number of consecutive failures
    Show answer & explanation

    Correct answers: B, C — Exponential interval, which multiplies the wait between consecutive attempts after a failure; Fixed interval, which keeps the wait between consecutive attempts at one constant length

    • A. A linear interval is a back-off style found in some libraries, but the action's Settings do not list it. The retry policy choices there are Default, None, Exponential interval and Fixed interval.
    • B. Exponential interval is a selectable retry policy type in the action's Settings. The wait grows after each failed attempt, which gives a throttled vendor API more time to recover.
    • C. Fixed interval is a selectable retry policy type in the action's Settings. It waits the same configured time between attempts, which suits a vendor whose throttling window is predictable.
    • D. Randomised jitter is a technique some SDKs and resilience libraries apply, but the action's Settings do not offer it as a retry policy type. The maker can only choose among the listed policies.
    • E. A circuit breaker belongs to resilience libraries and is not a retry policy type in the action's Settings. It stops calls after repeated failures instead of defining how retries are spaced.

    1.1: Plan an agent solution

    17.A company wants an internal agent that answers questions from SharePoint documents, but employees must only see content they already have permission to read. What should the plan include?

    1. A.Select Authenticate with Microsoft, add SharePoint knowledge through the maker's shared connection, and publish the agent to the whole tenant.
    2. B.Select Authenticate with Microsoft, add SharePoint knowledge so retrieval runs as the signed-in user, and publish the agent to the whole tenant.
    3. C.Select Authenticate with Microsoft, upload the SharePoint files as agent knowledge, and publish the agent with instructions limiting what it shares.
    4. D.Select No authentication, add SharePoint knowledge, and filter results by department with topic logic before publishing the agent to the tenant.
    Show answer & explanation

    Correct answer: B — Select Authenticate with Microsoft, add SharePoint knowledge so retrieval runs as the signed-in user, and publish the agent to the whole tenant.

    • A. Sign-in is required, but retrieval through the maker's shared connection runs with the maker's access. Every employee would see whatever the maker can read, not only what they are individually allowed to open.
    • B. Retrieval that runs as the signed-in user is security-trimmed by SharePoint, so each employee gets answers only from documents they can already open. That is the requirement.
    • C. Uploaded copies are stored as agent knowledge and no longer carry the SharePoint permissions of the originals. Instructions that limit what the agent shares are not access control, so restricted content can still surface.
    • D. Without authentication the agent has no user identity, so SharePoint cannot trim results per person. Department filtering in topics cannot reproduce document-level permissions, and restricted content can reach people not entitled to it.

    2: Integrate and extend agents in Copilot Studio

    2.1: Extend agents by using tools and enterprise data

    18.A maker adds this configuration in the Add MCP server page, but the connection to the internal inventory server cannot be established. Which line must change? ``` Name: Inventory MCP Description: Looks up warehouse stock and reorders items Server URL: http://inventory.contoso.internal/mcp Authentication: API key ```

    1. A.`Description: Looks up warehouse stock and reorders items`
    2. B.`Server URL: http://inventory.contoso.internal/mcp`
    3. C.`Authentication: API key`
    4. D.`Name: Inventory MCP`
    Show answer & explanation

    Correct answer: B — `Server URL: http://inventory.contoso.internal/mcp`

    • A. The description is free text that guides when the agent calls the server. A valid description is not a connection prerequisite.
    • B. The server URL must be an HTTPS endpoint reachable from Copilot Studio. The plain HTTP address is the faulty line.
    • C. API key is one of the supported authentication choices and fits a server that issues keys. The method is not what breaks the handshake here.
    • D. The name is a short label that helps the agent decide when to use the server. It does not block the connection.

    2.2: Configure multi-agent collaboration in Copilot Studio

    19.A finance team builds a Microsoft Copilot harness parent that delegates to two connected Copilot Studio agents. The administrator asks what to expect when makers test the parent and its connected agents during development. Which statement is accurate?

    1. A.Test conversations in the Preview tab are always free, because credits are charged only for conversations in published channels
    2. B.Credits are consumed only by the parent agent, because connected agents run on the parent's compute at no additional charge
    3. C.The harness uses usage-based billing, so using, building, testing and evaluating the agents can consume Copilot Credits
    4. D.Each connected agent must be licensed separately per end user before the parent can delegate to it in a test conversation
    Show answer & explanation

    Correct answer: C — The harness uses usage-based billing, so using, building, testing and evaluating the agents can consume Copilot Credits

    • A. Incorrect. Building, testing and evaluating can also consume credits on this harness.
    • B. Incorrect. Nothing in the guidance exempts connected agents from usage-based billing.
    • C. Correct. Agents on the Microsoft Copilot harness use usage-based billing, and testing and evaluation can consume Copilot Credits.
    • D. Incorrect. Delegation does not require a separate per-user licence for each connected agent.

    2.2: Configure multi-agent collaboration in Copilot Studio

    20.The orchestrator decides whether to delegate to a connected agent based on its name and ____, the user's message and context, and the parent's instructions.

    1. A.avatar image
    2. B.description
    3. C.publish date
    Show answer & explanation

    Correct answer: B — description

    • A. Incorrect. Icons aren't used for routing.
    • B. Correct. The description tells the orchestrator when the connected agent applies.
    • C. Incorrect. Dates are not routing inputs.

    2.1: Extend agents by using tools and enterprise data

    21.A company wants a connector tool to call a finance API with one service account the maker connected, not with each user's credentials. Which configuration achieves this?

    1. A.Keep the agent on an authenticated channel, set the tool's Credentials to use to Maker-provided credentials, then republish the agent.
    2. B.Keep the agent on an authenticated channel, share the maker's connection in Power Apps with Can use + share, then republish the agent.
    3. C.Keep the agent on an authenticated channel, set the tool's Credentials to use to User credentials after the maker signs in, then republish the agent.
    4. D.Switch the agent's channel to no authentication, let the tool use the maker's existing connection by default, then republish the agent.
    Show answer & explanation

    Correct answer: A — Keep the agent on an authenticated channel, set the tool's Credentials to use to Maker-provided credentials, then republish the agent.

    • A. Setting the tool's Credentials to use to Maker-provided credentials makes every call run under the maker's connected service account. This needs an authenticated channel and takes effect once the agent is republished.
    • B. Sharing a connection with Can use + share only lets other people reuse it in their own apps and flows. The tool still uses its default user credentials, so each user is still asked for their own sign-in.
    • C. The User credentials value makes the tool call the API with the signed-in user's own identity, and each user is prompted to connect. The maker signing in first does not change that, so the finance API is not called with one service account.
    • D. The tool does not fall back to the maker's connection by default; that behaviour comes from the Maker-provided credentials setting, which needs an authenticated channel. Removing authentication leaves that setting unusable and does not create a shared service account.

    2.1: Extend agents by using tools and enterprise data

    22.HR needs an agent to run the same onboarding process every time a manager asks: create an approval, wait for the response, transform the data and update two systems in a fixed sequence. Which tool type best fits?

    1. A.A workflow added to the agent as a tool, with the onboarding steps defined in advance and executed in the authored order whenever the agent invokes it.
    2. B.A Foundry IQ knowledge base loaded with the onboarding handbook, with the agent retrieving the documented steps and carrying them out as it reads them.
    3. C.A connector action added at agent level for the HR system, called by the agent once per request to carry out the onboarding in a single operation.
    4. D.An external MCP server exposing the HR and approval operations as tools, with the agent calling them in turn as the onboarding request progresses.
    Show answer & explanation

    Correct answer: A — A workflow added to the agent as a tool, with the onboarding steps defined in advance and executed in the authored order whenever the agent invokes it.

    • A. A workflow added as a tool encodes the steps and their order ahead of time, including pausing for a response and reshaping data. The agent runs the same flow on demand, which is what a fixed multi-system procedure needs.
    • B. Foundry IQ retrieves grounded content with citations from the handbook. Reading a procedure does not create approvals, wait for replies or write to systems, so nothing runs the steps.
    • C. A connector action performs one operation against one service. It cannot wait for an approval response, transform data between steps or update a second system, so the multi-step sequence is not covered.
    • D. MCP tools are exposed to the orchestrator, which decides which to call and when. Nothing fixes the order or enforces the wait for the approval response. A workflow defines the sequence explicitly.

    2.1: Extend agents by using tools and enterprise data

    23.Answers from a connected Foundry IQ knowledge base are weak. The agent's activity trace shows the Foundry IQ retrieval step ran but returned poorly ranked items. Which two actions are appropriate? Choose 2.(Select 2)

    1. A.Upload the highest-value documents as agent files in Copilot Studio, so the agent ranks them ahead of the Foundry IQ results.
    2. B.Work with the knowledge base owner in Foundry to adjust sources, retrieval instructions or ranking behavior, so better items come first.
    3. C.Remove the Foundry IQ connection in Copilot Studio and add it again, so the knowledge base's index is rebuilt with fresh ranking.
    4. D.Build a Copilot Studio topic that queries the search index over HTTP, so Power Fx can sort the returned items by relevance.
    5. E.Rewrite the Foundry IQ tool's name and description in Copilot Studio with details on its content, so orchestration sends it the right questions.
    Show answer & explanation

    Correct answers: B, E — Work with the knowledge base owner in Foundry to adjust sources, retrieval instructions or ranking behavior, so better items come first.; Rewrite the Foundry IQ tool's name and description in Copilot Studio with details on its content, so orchestration sends it the right questions.

    • A. Agent files are a separate knowledge source. They are not merged into the Foundry IQ result list and cannot re-rank it, so the retrieval step still returns the same poorly ranked items.
    • B. Ranking quality is determined by the knowledge base in Foundry, where its owner controls the knowledge sources, retrieval instructions and ranking behavior. The trace shows retrieval ran but ranked poorly, so that is where it is tuned.
    • C. The connection in Copilot Studio is only a reference to the knowledge base. Removing and re-adding it does not touch or rebuild the knowledge base or its index in Foundry, so ranking stays the same.
    • D. Querying the search index from a topic duplicates the knowledge base's retrieval and bypasses its centrally managed configuration. It leaves the ranking problem in Foundry unsolved and adds logic to maintain separately.
    • E. Orchestration relies on the tool's name and description to decide when to call it and what to send. Describing the knowledge base's content precisely helps the agent route suitable questions to it, which improves the answers.

    2.2: Configure multi-agent collaboration in Copilot Studio

    24.A parent agent delegates to a connected Benefits agent whose SharePoint tool uses Agent author authentication. During testing, a junior employee receives salary band documents that their own account cannot open. Which change fixes the cause?

    1. A.Publish the Benefits agent in a managed solution, so the SharePoint connection reference it relies on is locked and cannot be reused at runtime
    2. B.Add a rule to the parent's instructions telling the orchestrator to withhold salary band files from non-HR users, so it filters what it relays
    3. C.Switch the connected agent's SharePoint tool to User authentication, so each call runs under the signed-in employee's own access rights
    4. D.Remove the employee group's share on the Benefits agent, so the parent can no longer reach its tools on behalf of those users
    Show answer & explanation

    Correct answer: C — Switch the connected agent's SharePoint tool to User authentication, so each call runs under the signed-in employee's own access rights

    • A. Incorrect. A managed solution only protects components from being edited after import. The SharePoint tool still runs under the maker's identity at runtime, so the junior employee still receives files their own account cannot open.
    • B. Incorrect. Instructions guide the model's wording and behaviour but are not an access control. The tool still retrieves documents with the maker's permissions, so the exposure remains and relies on the model recognising salary content.
    • C. Correct. With Agent author authentication, the SharePoint tool uses the maker's access, which is broader than the junior employee's. User authentication makes each call run as the person chatting, so SharePoint returns only what that person may open.
    • D. Incorrect. Sharing controls who can chat with an agent directly, and the parent's delegation to the connected agent does not depend on it. The tool would keep running under the maker's identity, so the cause of the over-broad access stays.

    2.2: Configure multi-agent collaboration in Copilot Studio

    25.A remote A2A agent logs every request it receives from Copilot Studio. Which two items does a typical payload contain? (Choose two.)(Select 2)

    1. A.The latest user message without prior turns
    2. B.A connection reference for the parent agent
    3. C.A contextId that identifies the conversation
    4. D.Topic and variable definitions of the caller
    5. E.The version number of the exported solution
    6. F.The full chat history up to the new message
    Show answer & explanation

    Correct answers: C, F — A contextId that identifies the conversation; The full chat history up to the new message

    • A. Incorrect. Copilot Studio sends earlier turns along with the new message, so the remote agent is not limited to the latest utterance.
    • B. Incorrect. A connection reference is a Power Platform artifact that stays with the parent agent's environment. It is not part of the messages sent to a remote A2A agent.
    • C. Correct. Each A2A message carries a contextId that lets the remote agent tell which conversation the request belongs to.
    • D. Incorrect. Topic and variable definitions are configuration that stays in Copilot Studio. The remote agent receives conversation content, not that configuration.
    • E. Incorrect. A solution's version number is packaging metadata that belongs to the parent agent's environment and lifecycle management. It is not included in A2A requests.
    • F. Correct. Copilot Studio passes the conversation history as metadata with each request, so the remote agent sees prior turns as well as the latest one.

    2.2: Configure multi-agent collaboration in Copilot Studio

    26.A maker tests as an intended end user. Which three checks belong in that test? (Choose three.)(Select 3)

    1. A.Skip the restricted-data tests for tools that run through connections
    2. B.Confirm that User-authenticated tools return permitted data and refuse blocked data
    3. C.Share the child agent with the whole department, then test delegation as a member
    4. D.Share only the published parent with a test user and test delegation to the child
    5. E.Run the checks while signed in as the maker and compare results with the design
    6. F.Check that Agent author authentication tools return data fit for the parent's users
    Show answer & explanation

    Correct answers: D, B, F — Share only the published parent with a test user and test delegation to the child; Confirm that User-authenticated tools return permitted data and refuse blocked data; Check that Agent author authentication tools return data fit for the parent's users

    • A. Incorrect. Connection permissions still decide what a given user can see, so tools that run through connections need the restricted-data test like any other.
    • B. Correct. With User-authenticated tools, the end user's own permissions apply to each call. Requesting both permitted and blocked data shows whether access control holds for that user.
    • C. Incorrect. End users are given the published parent only, not the child. Sharing the child with a department gives the tester access real users lack, so a delegation failure for them would go unnoticed.
    • D. Correct. End users receive only the published parent. A test user with just that share shows whether delegation to the child works without extra permissions.
    • E. Incorrect. A maker has broader access than intended end users, so a test signed in as the maker hides permission differences. The checks have to run as an end user.
    • F. Correct. Tools using Agent author authentication run with the author's credentials, so anyone who can use the parent may see what those credentials return. The returned data has to suit that audience.

    3: Test and manage agents

    3.2: Implement application lifecycle management (ALM) for agents in Copilot Studio

    27.Before exporting, select Remove from this solution on an environment variable's ___ so the development value stays behind and the target environment must supply its own.

    1. A.Current Value
    2. B.Display name
    3. C.Default Value
    Show answer & explanation

    Correct answer: A — Current Value

    • A. Removing the current value keeps it in development but leaves it out of the exported package.
    • B. The display name is only a label and has no effect on which value is exported.
    • C. The default value is part of the definition and is only a fallback when no current value exists.

    3.1: Evaluate agent performance

    28.A compliance test case requires that the agent's answer mentions both 'cooling-off period' and 'written notice'. Missing either phrase must fail the case. Reading the configuration below, which setting is at fault? ``` Test method: Keyword match Match mode: Any Keywords: cooling-off period; written notice ```

    1. A.Keywords entry `cooling-off period`, because hyphenated terms are rejected and must be typed with spaces instead.
    2. B.Keywords entry `written notice`, because keyword phrases must be capitalised to be compared with the response text.
    3. C.Match mode `Any`, which passes a case when only one of the two phrases appears, so it should be `All`.
    4. D.Test method `Keyword match`, which cannot check phrases of more than one word and must be replaced with Exact match.
    Show answer & explanation

    Correct answer: C — Match mode `Any`, which passes a case when only one of the two phrases appears, so it should be `All`.

    • A. Incorrect. Hyphens are fine in keywords; the problem is how many phrases must match, not the characters used.
    • B. Incorrect. Capitalisation is not what decides this scenario; the match mode controls whether one or all phrases are required.
    • C. Correct. Requiring every phrase means the mode must be All; Any lets an answer with a single phrase pass.
    • D. Incorrect. Keyword match accepts words and multi-word phrases; Exact match would demand the whole answer be identical.

    3.2: Implement application lifecycle management (ALM) for agents in Copilot Studio

    29.Contoso built a support agent directly in its development environment and never created a solution. The team now needs to move the agent to a test environment. What should they do first?

    1. A.Create a custom solution with the team's own publisher in development, then add the existing agent to it via Add existing > Agent.
    2. B.Create a custom solution in the test environment, then pull the support agent into it from development with Add existing > Agent.
    3. C.Start a Power Platform pipeline from the Default Solution in development, then promote the support agent straight to the test environment.
    4. D.Export the Default Solution from development as a package, then import it into the test environment to bring the agent over.
    Show answer & explanation

    Correct answer: A — Create a custom solution with the team's own publisher in development, then add the existing agent to it via Add existing > Agent.

    • A. Agents travel between environments inside solutions, and this agent sits only in the Default Solution. A custom solution in development, with Add existing > Agent, gives it a deployable container.
    • B. Add existing lists only components that already exist in the current environment. A solution created in test cannot see the agent that lives in development, so nothing can be added to it.
    • C. Pipelines deploy a custom solution from the development environment, and the Default Solution cannot be their source. The agent must first be placed in a custom solution.
    • D. The Default Solution is the environment-wide container for every component, not a deployment unit for one agent. The agent needs its own custom solution to be exported and imported cleanly.

    3.2: Implement application lifecycle management (ALM) for agents in Copilot Studio

    30.An admin wants a Power Automate flow in the pipelines host to start whenever a deployment is requested. How should the flow's trigger be configured?

    1. A.Use the Copilot Studio When an agent calls this flow trigger, and register the flow as a tool on the agent that belongs to the pipelines host.
    2. B.Use a Recurrence trigger on a five-minute interval, and list the pipeline run history through the Power Platform for Admins connector.
    3. C.Use Dataverse When an action is performed, catalog Microsoft Dataverse Common, category Power Platform Pipelines, action OnDeploymentRequested.
    4. D.Use the Dataverse When a row is modified trigger on the Solution table in the development environment, with Scope set to Organization.
    Show answer & explanation

    Correct answer: C — Use Dataverse When an action is performed, catalog Microsoft Dataverse Common, category Power Platform Pipelines, action OnDeploymentRequested.

    • A. The When an agent calls this flow trigger only fires when a Copilot Studio agent invokes the flow as a tool. A pipeline deployment request does not call an agent, so the flow would never start from that event.
    • B. A Recurrence trigger starts the flow on a schedule, not in response to the deployment request. Reading run history afterwards delays the start by up to the polling interval and consumes runs even when nothing was requested.
    • C. Pipeline events surface as Dataverse custom actions under that catalog and category.
    • D. Changes to Solution rows are not a signal that a deployment was requested, so this trigger would fire on unrelated edits and not on the request itself. The flow also has to live in the pipelines host.

    3.2: Implement application lifecycle management (ALM) for agents in Copilot Studio

    31.An approver rejects a deployment in a flow that uses the Perform an unbound action step. The maker must see why it was rejected and the deployment must stop. What should the flow set?

    1. A.Set the step status to 30 and write the reason into the properties field of the deployment stage run record in Dataverse.
    2. B.Set the step status to 10 and write the reason into the properties field of the deployment stage run record in Dataverse.
    3. C.Set the step status to 30 and write the reason into the comments field of the deployment stage run record in Dataverse.
    4. D.Set the step status to 20 and write the reason into the comments field of the deployment stage run record in Dataverse.
    Show answer & explanation

    Correct answer: C — Set the step status to 30 and write the reason into the comments field of the deployment stage run record in Dataverse.

    • A. Status 30 does reject the step and stop the deployment, but the properties field is not shown to makers, so they would not see why it was rejected.
    • B. Status 10 is the pending state, so the step is not rejected and the deployment keeps waiting. The properties field is also not shown to makers.
    • C. Status 30 marks the step as rejected, which fails the deployment. The comments field is shown to the maker in the pipeline run history.
    • D. Status 20 marks the step as completed, so the deployment would proceed. The comment would be visible, but it would contradict the status and the deployment would not stop.

    3.1: Evaluate agent performance

    32.A support agent must answer a refund question with the exact policy sentence 'Refunds are issued within 14 days of receiving the returned item.' Legal wants wording and sentence structure to match closely, with a tunable tolerance. Which test method fits best?

    1. A.Compare meaning, which judges whether the agent answer expresses the same idea as the expected response sentence, however it is phrased.
    2. B.Text similarity, which scores the agent answer against the expected response sentence using cosine similarity over the text of both.
    3. C.General quality, which rates relevance, groundedness and completeness of the agent answer, judged against the retrieved policy content.
    4. D.Tool use, which verifies the agent called the knowledge source holding the refund policy before it produced the answer.
    Show answer & explanation

    Correct answer: B — Text similarity, which scores the agent answer against the expected response sentence using cosine similarity over the text of both.

    • A. Incorrect. Compare meaning rewards an answer that carries the same idea, so a reworded refund statement can pass. Legal needs the wording and sentence structure to stay close to the policy sentence.
    • B. Correct. Text similarity compares the answer's text with the expected response using cosine similarity, so near-verbatim wording scores high. It also has a pass score the maker can tune, which gives the adjustable tolerance.
    • C. Incorrect. General quality rates relevance, groundedness and completeness without comparing against an expected sentence. A well-grounded answer in different words would still score well, so it does not enforce the required wording.
    • D. Incorrect. Tool use only checks which tools or knowledge sources were invoked. It says nothing about the wording of the final response, so it cannot enforce the exact policy sentence.

    3.1: Evaluate agent performance

    33.An order-lookup agent returns a status code such as 'SHIPPED-EU' for each question. The tester needs a binary pass only when the whole answer equals the expected code, character for character. Which test method should be added?

    1. A.Exact match with the expected code in the test set, which grades the answer pass or fail on whether the text is identical to that code.
    2. B.Custom with Valid and Invalid labels and written instructions, which grades the answer pass or fail on how a model applies those labels.
    3. C.Compare meaning with a pass score of 100, which grades the answer pass or fail on whether it conveys the same intent as that code.
    4. D.Keyword match with the status code as the keyword, which grades the answer pass or fail on whether that code appears in the text.
    Show answer & explanation

    Correct answer: A — Exact match with the expected code in the test set, which grades the answer pass or fail on whether the text is identical to that code.

    • A. Correct. Exact match gives a binary result and passes only when the whole response is identical to the expected value, which is what a short status code such as SHIPPED-EU needs.
    • B. Incorrect. A custom grader is driven by evaluation instructions and labels, so a model decides which label fits. That is a judgment call, not a deterministic character-for-character comparison with the expected code.
    • C. Incorrect. Compare meaning scores whether the response expresses the same intent as the expected one, even with a high pass score. A reworded answer that never prints the exact code can still pass, which breaks the equality requirement.
    • D. Incorrect. Keyword match only checks that the keyword occurs in the response. An answer that wraps the code in extra text passes, although the whole answer must equal the expected code.

    3.1: Evaluate agent performance

    34.A maker opens a failed test case in the evaluation results. Which THREE items does the case detail view provide to help diagnose the failure? (Choose 3.)(Select 3)

    1. A.A Show activity map view tracing the agent's inputs, decisions and outputs for that test case.
    2. B.A change log of who edited the agent's instructions and topics between the previous run and this run.
    3. C.The knowledge sources, topics and tools the agent drew on for its response, each opened from the case.
    4. D.The reasoning behind the result for the chosen test method, shown next to the expected and actual responses.
    5. E.A downloadable Application Insights trace of the Dataverse calls made while the agent handled that case.
    Show answer & explanation

    Correct answers: A, C, D — A Show activity map view tracing the agent's inputs, decisions and outputs for that test case.; The knowledge sources, topics and tools the agent drew on for its response, each opened from the case.; The reasoning behind the result for the chosen test method, shown next to the expected and actual responses.

    • A. Correct. The Show activity map view lets the maker follow the agent's flow of inputs, decisions and outputs for the selected test case.
    • B. Incorrect. The case detail view reports on the test case itself: the responses, the result and how the agent produced them. It does not include an edit history of the agent's instructions or topics, so it cannot show what changed between runs.
    • C. Correct. The detail view lists the knowledge sources, topics and tools the agent used for the response, and the maker can open each one to inspect it.
    • D. Correct. The detail view explains why the case passed or failed.
    • E. Incorrect. The case detail view has no export of Application Insights telemetry or Dataverse calls. Diagnosis from the evaluation results relies on the activity map, the resources used and the result reasoning.

    3.1: Evaluate agent performance

    35.The Evaluate agent button appears in the Review panel as a warning. Which TWO conditions produce that warning? (Choose 2.)(Select 2)

    1. A.The agent relies on a Microsoft Fabric data agent, so publishing is held until an evaluation has passed.
    2. B.The agent's model was changed after its latest evaluation, so the earlier results predate the new model.
    3. C.The agent has not been through an evaluation yet, so there are no results on file to review for it.
    4. D.The latest evaluation pass rate is below 80 percent, so the agent is switched off in its published channels.
    5. E.The agent has more than 100 topics, so evaluation is required before another topic can be added.
    Show answer & explanation

    Correct answers: B, C — The agent's model was changed after its latest evaluation, so the earlier results predate the new model.; The agent has not been through an evaluation yet, so there are no results on file to review for it.

    • A. Incorrect. Using a Fabric data agent does not hold up publishing. The warning reflects the agent's evaluation status, not which knowledge sources it uses.
    • B. Correct. Changing the agent's model after the latest evaluation leaves those results describing a different model, so the Review panel shows a warning.
    • C. Correct. An agent that has never been evaluated has no results on file, and the Review panel shows a warning for that.
    • D. Incorrect. A low pass rate does not switch an agent off in its channels. The warning concerns missing or outdated evaluation results.
    • E. Incorrect. The number of topics does not trigger an evaluation warning, and there is no rule requiring evaluation before another topic is added.

    Want the full experience?

    These are just samples. Practice the full Microsoft Power Platform AI Agent Builder Associate (AB-620) - December 18 onwards question bank in quiz mode — free, no signup, with domain practice and exam simulation.