Free Salesforce Certified Agentforce Specialist practice test: 30 questions with answers

Topics define the scope of what the agent handles, and actions expose approved steps such as flows, which keeps behaviour bounded. Unrestricted access removes the control the team wants, a text-only prompt cannot perform the return, and a single Apex method with no topics gives no structure for the agent.

Explicit instructions and guardrails in the topic, verified by testing with adversarial prompts, address the behaviour where it is defined. Deleting pricing data breaks legitimate uses, raising temperature makes outputs less predictable rather than more cautious, and removing all actions blocks valid tasks.

An escalation action routes the conversation to a queue with the context, so the human agent does not start from zero. A canned message forces the customer to repeat everything, automatic closure leaves the customer without help, and a weekly email is far too slow.

Putting the rules in a Flow or Apex action keeps them in one maintainable place that the agent calls. Copying conditions into every topic, hard-coding them in a greeting and an unmaintained notes field all create drift that a quarterly change would break.

Grounding in a knowledge source retrieves relevant documents and lets the reply cite them, which supports verifiable answers. Answering from general training data invents or misremembers policies, copying everything into instructions is unmanageable, and turning off reasoning does not retrieve policies.

Toxicity detection in the Einstein Trust Layer and instructions on tone for the channel address inappropriate language directly. Disabling the agent for the channel is unnecessary, ignoring tone is a poor customer experience, and one-word answers are unhelpful.

A Flow action with inputs creates the case with the right fields populated reliably. A blank screen makes the customer re-enter details, copying from chat is error-prone, and a free-text email requires manual entry later.

Passing the authenticated record context through the session lets the agent act for the right customer without repeated prompts and without exposing data. Repeated identifiers add friction, public account numbers leak data, and disabling authentication exposes every customer's records.

Packaging the configuration in source control and deploying through a reviewed pipeline with a rollback path makes the change repeatable and reversible. Manual recreation and direct production edits are error-prone and unrecorded, and never changing the agent prevents needed fixes.

Reviewing logs and traces alongside recent configuration changes finds the cause from evidence. Deleting logs removes the evidence, a full redesign stops service without a diagnosis, and switching models without checking causes may fix nothing.

A staged rollout to a small group with monitoring limits the impact of a faulty change and allows expansion when results are acceptable. An untested switch-over relies on complaints, codes add friction without monitoring, and disabling logging hides problems.

The setup audit trail and version history record each change with the user and time, which is the evidence the company needs. A shared login removes individual attribution, screenshots are incomplete, and a note does not show who changed what.

Realistic utterances including ambiguous requests and reviewed results against expected outcomes reveal real weaknesses. Happy-path-only tests miss failures, skipping minor changes ignores real effects, and production-only testing exposes customers.

Packaging and deploying the configuration, then verifying it, is repeatable and checkable. Manual recreation is error-prone, copying a database would also move customer data, and rebuilding from memory is unreliable.

A controlled comparison of old and new versions on the same scenarios, with recorded metrics, gives reliable evidence. Impressions from two conversations, counts of changes and page views do not show improvement.

The safest fix is to keep sensitive fields out of the prompt and use masking, so the model never receives them. A note asking the model to hide numbers is unreliable, passing every field increases exposure, and publishing drafts is a serious leak.

Explicit structure, order, length and an example make the output predictable. Leaving structure open increases variation, repeating a sentence adds noise without clarity, and changing the model each time adds inconsistency.

Listing the allowed values, asking for exactly one, and providing a fallback constrains the output to valid categories. Inventing categories, removing the list and choosing by length all produce values the business cannot use.

A conditional section includes the sentence only when there is data, which avoids awkward text without inventing values. Letting the model guess invents facts, removing the field loses information, and forcing users to type placeholder values adds work and errors.

Grounding in verified data and restricting answers to provided sources reduces invented features, and sampled review catches what remains. Asking for more confidence encourages overstatement, removing data invites invention, and publishing unreviewed answers exposes customers.

Placing the question first and including only relevant features focuses the model on what the customer asked. Random full catalogues add noise, reading the list aloud is not useful, and removing the question stops the answer from addressing the request.

An explicit length limit in the prompt, checked against samples, produces the required length. Leaving length to the model is unpredictable, asking for long text and letting the app cut it off breaks sentences, and long summaries do not suit mobile use.

Identity resolution matches records across sources into one unified profile that the agent can use. A dashboard does not unify records, a yearly spreadsheet is stale, and turning off source systems destroys the data.

A retriever over an indexed knowledge source returns current content as the index refreshes. A yearly copy in instructions is stale within weeks, remembering articles does not update them, and sending customers to search manually abandons the agent's purpose.

Data governance controls and data use purposes enforce which data can be used for which purpose. A documentation note is advisory, universal access breaks the limits, and copying to a public dataset removes the controls.

A calculated insight computes metrics from unified data and refreshes on a schedule, keeping the score current. Manual fields go stale, sorted names do not give a value, and screenshots are not a data source the agent can use.

Checking the data stream or connector confirms whether the update is being ingested, which is where the gap is. Topic wording, field colour and agent count do not control whether source updates reach Data Cloud.

Consistent mapping, reviewed identity resolution and monitoring of ingestion all improve data quality for the agent. Inconsistent field names, and keeping unverified old records, make answers less reliable.

The unified individual profile in the unified data model, built by identity resolution, combines purchase and web data for segmentation. Spreadsheets are manual, the Contact record lacks web activity, and a dashboard of past counts is not a profile source.

A clear handoff contract and a documented interface keep the agents interoperable while each keeps its responsibilities. Full access to another agent's data, merging agents and runtime changes to each other's instructions break separation of responsibility and create unpredictable behaviour.