Salesforce Certified Agentforce Specialist Practice Exam

What the Salesforce Certified Agentforce Specialist exam covers

• AI Agents278 questions
• Prompt Engineering130 questions
• Data 360 Fundamentals144 questions
• Testing, Deployment, and Maintenance105 questions
• Multi-Agent Orchestration43 questions
• Governance and Observability64 questions

Free Salesforce Certified Agentforce Specialist practice test questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 764.

1. Question 1AI Agents

   A Salesforce admin is configuring a brand new Agentforce agent for the first time. Which statement correctly describes how the agent is structured in Agent Builder?

   • AActions attach to the agent directly, bypassing topics
   • BTopics contain agents; each agent has one instruction set
   • CAgent contains topics; each holds instructions and actionsCorrect
   • DInstructions contain topics; topics hold the agent record
   ✓ Correct answer: C

   The correct model is agent to topics to instructions/actions: an Agentforce agent is the top-level container that holds one or more topics, and each individual topic carries its own instructions defining scope plus the actions it is allowed to invoke. This layered structure is what lets the Atlas Reasoning Engine take a broad agent purpose and break it into narrow, classifiable jobs - when a message comes in, the engine matches it to the best-fitting topic, then uses that topic's instructions and actions to build a grounded reply. For a brand-new agent, this is the very first design decision an admin makes: how to split the intended work into topics before writing instructions or wiring up actions. The other options each invert or scramble this hierarchy in a different way. Remember the order: agent owns topics, topics own instructions and actions, never the reverse.

   Why the other options are wrong

   • AActions are always scoped to a topic, not attached to the agent directly; there is no mechanism that lets an action bypass topics entirely.
   • BThis reverses the real hierarchy: an agent owns many topics, and each topic carries its own instruction set, not one instruction set shared and owned by all topics.
   • DInstructions belong to a topic and topics belong to the agent; instructions do not contain topics, and topics do not hold the agent record.
2. Question 2AI Agents

   A software vendor's agent must both look up a customer's current subscription tier and, if eligible, generate an upsell message describing a relevant add-on. What best describes the recommended action sequencing?

   • AThe lookup runs first and guides the upsell message generationCorrect
   • BThe upsell message is always generated first, regardless of tier
   • CBoth actions run in a completely random order
   • DOnly the lookup runs; a rep writes the message manually
   ✓ Correct answer: A

   This is a case of action sequencing and chaining: the lookup runs first, and its result, the customer's actual subscription tier and eligibility, becomes the grounding context the Prompt Template action uses to generate an upsell message that is accurate and relevant rather than generic. Generating the message before the tier is known risks recommending an add-on the customer is not even eligible for, sequencing is driven by this kind of logical dependency rather than being random, and handing message drafting to a human bypasses the automation the Prompt Template action exists to provide. Takeaway: when one action's output should shape another's content, sequence the data-gathering action first.

   Why the other options are wrong

   • BGenerating the upsell message before checking eligibility risks recommending an add-on the customer does not actually qualify for.
   • CAction sequencing follows logical dependency between actions, not a random order.
   • DHanding message drafting to a human bypasses the automation the Prompt Template action is meant to provide.
3. Question 3AI Agents

   A bank's compliance team asks why a Service Agent cannot surface a customer's full account balance history when the underlying record is restricted by a sharing rule to a specific team. What is the explanation?

   • AAgentforce ignores sharing rules for efficiency
   • BSharing rules only apply to Experience Cloud, not agents
   • CThe agent needs a separate Trust Layer setting to respect sharing
   • DRecord-level security is enforced during execution, like any userCorrect
   ✓ Correct answer: D

   The agent's actions run as its Agent User, and any sharing rule that restricts a record, such as an account balance history limited to a specific team, applies to that user exactly as it would to a human logging in directly, which explains why the agent cannot surface data the Agent User is not shared into. The key concept is that Agentforce introduces no separate, bypassable security layer; it operates within the standard Salesforce sharing model, so whatever restricts a record from a regular user restricts it from the agent too. This is precisely why the compliance team's expectation is being met correctly rather than indicating a bug: the restriction is working as intended. Claiming Agentforce ignores sharing rules for efficiency misstates how the platform actually enforces security, saying sharing rules apply only to Experience Cloud ignores that they apply broadly across all Salesforce access including agent execution, and there is no separate Trust Layer toggle needed since standard sharing enforcement applies by default without extra configuration. Takeaway: agent execution is bound by the same sharing rules as any Salesforce user, with no bypass.

   Why the other options are wrong

   • AAgentforce does not ignore sharing rules for efficiency; they are actively and consistently enforced during execution.
   • BSharing rules apply broadly across all Salesforce access, including agent execution, not only to Experience Cloud users.
   • CThere is no separate Trust Layer setting required to enforce standard record sharing; it applies by default without extra configuration.
4. Question 4AI Agents

   A telecom agent asks the customer to choose a service plan tier in an early turn, then references that choice in a later turn to calculate a quote, all within the same ongoing conversation. Which variable type is designed to hold this value between turns?

   • AAn output variable
   • BAn input variable
   • CA conversation variableCorrect
   • DA toxicity filter
   ✓ Correct answer: C

   A conversation variable is designed to hold the chosen plan tier between the early turn where the customer selects it and the later turn where the agent needs it again to calculate the quote, all within one ongoing conversation. This avoids the agent asking the customer to restate an earlier choice, which is exactly the within-session memory problem conversation variables solve. An output variable instead hands a value back out to the host application rather than persisting it for the agent's own later use, an input variable is populated once from the calling context at launch rather than from a choice made mid-conversation, and there is no such element as a toxicity filter for holding conversational state, toxicity detection is a Trust Layer content safeguard entirely unrelated to remembering values. The takeaway: choices the customer makes partway through a conversation, needed again later in that same conversation, belong in a conversation variable.

   Why the other options are wrong

   • AAn output variable is used to hand a value back out to the host application, not to persist it across turns for the agent's own later use.
   • BAn input variable is populated once from the calling context at launch, not from a choice the customer makes mid-conversation.
   • DThere is no such element as a toxicity filter for holding conversational state; toxicity detection is a Trust Layer content safeguard, unrelated to variables.
5. Question 5Data 360 Fundamentals

   A logistics company's agent needs to reference both a CRM Shipment custom object and a Data Cloud data graph relating Shipment, Stop, and Package. Where are both added so the agent can ground on them?

   • AOnly the CRM object needs the Library
   • BBoth need to be added to the Data LibraryCorrect
   • COnly in the Prompt Builder template
   • DOnly in the Testing Center test set
   ✓ Correct answer: B

   Regardless of whether a source is a native Salesforce object like the Shipment custom object, or a Data Cloud data graph relating Shipment, Stop, and Package, both are added to the agent through the same mechanism, the Data Library in Agent Builder. Neither structured CRM records nor a Data Cloud-native relationship structure gets any special-case shortcut; the Library is the single connection point for all grounding sources. This logistics company's agent needs both, so the admin simply adds both there rather than treating one as automatic or routing either through an unrelated feature. Prompt Builder templates and Testing Center test sets serve later stages, prompt authoring and behavior validation respectively, not source connection. The rule to remember: whatever kind of content an agent needs to ground on, it gets added in one place.

   Why the other options are wrong

   • AA data graph is never automatically available to an agent; like a CRM object, it must be explicitly added to the Data Library before any topic can use it.
   • CPrompt Builder is where prompt templates reference already-connected data through merge fields; it is not where grounding sources themselves get added to an agent.
   • DThe Testing Center is used to validate an agent's behavior after its grounding sources are already configured, not to connect a CRM object or data graph.
6. Question 6Data 360 Fundamentals

   A law firm configuring its research agent's retriever wants to be clear with stakeholders about a governance limit: retrieval cannot be used to work around Salesforce's sharing model. Which statement accurately reflects this limit?

   • AEnsemble retrievers are exempt from record-level security
   • BField-level security applies only to prompt templates
   • CRetrieved content is still subject to the user's own accessCorrect
   • DSecurity enforcement only applies for keyword search
   ✓ Correct answer: C

   The governance limit being described is precisely that retrieval is not a special path around the sharing model, whatever a retriever surfaces still has to pass through the same record and field-level access checks that apply to the running user everywhere else in the org. The key concept is consistency of enforcement across access paths, a retriever is just another way data can be reached, not an exception to how it is protected. This matters for a research agent because it may search across a broad range of indexed matter, and stakeholders need confidence that broad searchability does not translate into broad, unauthorized visibility. The other options claim exemptions for a retriever type, a narrower scope for field-level security, or a dependency on search index type, none of which describes how enforcement actually works. Remember: no retriever type or index type is exempt from access control.

   Why the other options are wrong

   • ANo retriever type, including ensemble retrievers, is exempt from record-level security; enforcement applies regardless of retriever configuration.
   • BField-level security applies broadly to data access across the platform, including content surfaced through agent topics, not only within prompt templates.
   • DSecurity enforcement is not tied to a specific search index type; it applies the same way whether the index is keyword, vector, or hybrid.
7. Question 7Testing, Deployment, and Maintenance

   A university's admissions agent, while in sandbox testing, occasionally answers questions about tuition using outdated figures. What is the most likely explanation the team should investigate?

   • AThe grounding source (Knowledge, Data Cloud, or record data) has stale content in the test environmentCorrect
   • BThe Testing Center itself caches all responses permanently across every environment
   • CThe agent's topic was deployed without a domain assignment
   • DThe sandbox license has expired
   ✓ Correct answer: A

   Occasional outdated tuition figures point to the grounding source itself, Knowledge, Data Cloud, or record data, holding stale content in the test environment, since a grounded response is only as accurate as the material it pulls from. The fix is refreshing or correcting that source content in the sandbox, exactly the kind of gap sandbox testing is meant to catch before production. The Testing Center does not permanently cache conversational responses across environments, so that is not the explanation. Domain assignment refers to how exam content is organized into topic areas for study purposes and has nothing to do with grounding freshness in a live agent. Sandbox licenses do not expire in a way that would cause stale content to appear. Investigating and refreshing the actual source data is the right next step.

   Why the other options are wrong

   • BThe Testing Center does not permanently cache conversational responses across environments; that is not a real behavior of the tool.
   • CDomain assignment is a way exam content is organized into topic areas and has no connection to grounding freshness in a live agent.
   • DSandbox licenses do not expire in a way that would cause previously fine content to suddenly appear stale.
8. Question 8Testing, Deployment, and Maintenance

   At a car dealership finance arm, two admins both want to edit the live 'Loan Application Status' topic in the same week. What practice best reduces the risk of one admin's change unexpectedly overwriting the other's work?

   • ACoordinate changes through a shared sandbox and controlled deployment process with visible change historyCorrect
   • BLet each admin edit the live production topic independently at the same time
   • CHave one admin disable the topic while the other edits it, with no communication at all
   • DAvoid documenting either change so there is no confusion about who did what
   ✓ Correct answer: A

   When two admins want to edit the same live topic in the same week, the risk of one change silently overwriting the other is reduced by coordinating through a shared sandbox and a controlled deployment process with visible change history, giving both admins a single, tracked path to production instead of two independent, colliding edits. The key concept is that production topics are a shared resource, so uncoordinated simultaneous changes create a real conflict risk that only visibility and a single controlled promotion path can prevent. For the car dealership finance arm's Loan Application Status topic, this coordination is what lets both admins contribute safely in the same week. B is precisely the scenario that risks an overwrite, C disrupts customers and still does not resolve the coordination problem, and D makes conflicting changes harder, not easier, to untangle. Takeaway: coordinate concurrent edits to a shared live topic through one controlled process with visible change history.

   Why the other options are wrong

   • BLetting both admins edit the same live production topic independently and simultaneously is exactly the scenario that risks one change overwriting the other.
   • CDisabling the topic while one admin edits it, with no communication to the other, disrupts customers and still fails to resolve the coordination problem.
   • DAvoiding documentation of either change makes it harder, not easier, to understand who changed what or resolve a conflict.
9. Question 9Prompt Engineering

   A Service Agent topic must summarize a customer's three most recent open Cases when the customer asks what is happening with their issues. The Cases are child records of the Contact being actioned. Which grounding technique fits without writing Flow or Apex?

   • AMerge-field grounding on the Contact record only
   • BData Cloud retriever grounding over files
   • CRelated-list grounding on the Contact's CasesCorrect
   • DApex-based grounding with a custom query
   ✓ Correct answer: C

   Related-list grounding is the declarative technique built for surfacing a set of connected child records, such as the Cases related to a Contact, so the prompt can reference or summarize several of them at once. Because the requirement explicitly rules out writing Flow or Apex, this points directly at related-list grounding rather than a custom-built alternative that could technically achieve the same result with more effort. The key limitation to recognize with a single merge field is that it can only resolve one value at a time, it cannot enumerate multiple Case records, which is precisely why related-list grounding exists as its own distinct technique for this exact use case.

   Why the other options are wrong

   • AA single merge field can only resolve one value at a time, it cannot enumerate or summarize several related Case records at once.
   • BThe Cases here are structured CRM records already in the org, not unstructured content requiring a Data Cloud retriever's semantic search.
   • DApex-based grounding with a custom query would work but is unnecessary effort here, since related-list grounding already exposes the needed child records declaratively without any code.
10. Question 10Governance and ObservabilitySelect all that apply

    A Calderon Health session ended with the agent apologising and closing the chat, and the team is opening the observability record to work out what happened. Which two pieces of information can they get for that session? (Choose TWO)

    • AThe error status and error type recorded against the sessionCorrect
    • BThe Apex stack trace for every action that the agent invoked
    • CThe full system prompt text sent to the model on each turn
    • DThe knowledge sources the agent used during that sessionCorrect
    ✓ Correct answer: A, D

    Two facts about a failed session are available directly: the error status with its type, which says whether the cause was a system problem, an API failure, or agent logic, and the knowledge sources the agent drew on, which can also be added as a field in the Sessions and Intents table. Between them they usually separate an integration failure from a grounding problem, which is the first fork in any investigation. An Apex stack trace is developer debugging output produced by Apex execution and captured in debug logs, not something the session record presents. The full system prompt text sent to the model is likewise not exposed as session observability data, and expecting it confuses a prompt-engineering artifact with a monitoring record. The takeaway is that observability answers what happened and what grounded it, while low-level code debugging stays in developer tooling. Release note: the error status and error type columns are documented in the Summer 26 release notes, available the week of July 6, 2026, and knowledge source visibility comes from the Summer 26 Refined Agent Analytics entry, whereas the exam guide states that exam questions align to the Spring 26 release. Learn it as current product behaviour rather than assuming it is in scope for an exam whose stated baseline is Spring 26.

    Why the other options are wrong

    • BApex stack traces are developer debug output captured in debug logs rather than data presented on a session.
    • CThe raw prompt text sent to the model is a prompt-engineering artifact, not information the session record exposes.

Who this Salesforce Certified Agentforce Specialist practice exam is for

This practice set is for anyone preparing for the Salesforce Certified Agentforce Specialist exam - from first-time candidates building a foundation to experienced Salesforce practitioners doing a final review before test day. If you learn best by working through realistic questions and reading why each answer is right or wrong, it is built for you.

How to use this Salesforce Certified Agentforce Specialist practice exam

1. Start with the free sample questions above to gauge your current baseline.
2. Read the full explanation on every question, including why each wrong option is wrong.
3. Track your weak domains and focus your study where you are losing the most marks.
4. Once you are scoring consistently well, take a timed, full-length mock exam.
5. Use your readiness score to decide when you are ready to book the real Salesforce Certified Agentforce Specialist exam.

Related Salesforce resources

• Salesforce Certified Agentforce Specialist study guideKey concepts
• Salesforce practice examsAll Salesforce
• Certification pathWhere this fits
• Certification exam guides & tipsBlog
• Plans & pricingFree & paid
• How these questions are written and reviewedMethodology
• Report a problem with a questionCorrections
• Salesforce Certified Business Analyst practice examRelated
• Salesforce Certified Data 360 Consultant practice examRelated
• Salesforce Certified Platform Administrator (ADM-201) practice examRelated

Salesforce Certified Agentforce Specialist practice exam FAQ