# Free Salesforce Agentforce Specialist Practice Questions

*Last updated September 2026 · https://www.certifyforce.com/salesforce-agentforce-specialist-practice-exam/free-questions*

18 free Agentforce Specialist practice questions from the CertifyForce bank of 305 reviewed questions, spread across the 6 official exam sections in proportion to their weight. Each has the correct answer, an explanation, and the official documentation it was verified against.

## Prompt Engineering (20%)

### Question 1

Cloud Kicks needs a single prompt template that combines data from a custom Reservation object and a related Experience object, and the resulting text will be dropped into a PDF, a newsletter, and an agent action. Which prompt template type should the Agentforce Specialist choose?

- A. Flex, because it has no predefined entry point and can reference multiple object resources.
- B. Field generation, because it can be assigned to several fields on the record layout.
- C. Record summary, because it can pull from multiple related lists on one record page.
- D. Sales email, because it applies rich formatting suited for newsletters.

**Correct answer:** A. Flex, because it has no predefined entry point and can reference multiple object resources.

**Explanation:** Flex templates are the general-purpose type: they accept multiple object resources plus free-text inputs, and because they have no predefined entry point they can be called from Flow, Apex, the REST API, or Agentforce and reused for many outputs. Field generation is bound to one writable field on one object and is triggered from the record page, so it cannot feed a PDF or an agent action. Record summary produces a summary for a single record's UI context rather than a reusable multi-object output. Sales email is scoped to the email composer entry point regardless of its formatting.

**Sources:** [Create a Flex Prompt Template with Einstein AI](https://trailhead.salesforce.com/content/learn/projects/quick-start-create-prompt-builder-flex-template/build-a-flex-prompt-template), [Maximizing Generative AI with Effective Prompts](https://trailhead.salesforce.com/content/learn/modules/prompt-builder-basics/get-to-know-prompt-builder)

### Question 2

Cumulus Financial wants a generated onboarding recap produced automatically whenever an Opportunity reaches Closed Won, with the text stored on a related record. An active flex prompt template already produces the recap. How should the Agentforce Specialist execute the template?

- A. Add the template as a generative field on the Opportunity Lightning record page so users click to generate it.
- B. Ask reps to run the template from the Einstein sales email composer and paste the result.
- C. Create a formula field that references the prompt template API name.
- D. Call the template from a record-triggered flow using the prompt template flow action and store the returned text.

**Correct answer:** D. Call the template from a record-triggered flow using the prompt template flow action and store the returned text.

**Explanation:** Flex templates are designed to be invoked programmatically, and a record-triggered flow with the prompt template action runs automatically on the Closed Won update and returns the generated text for the flow to save. A generative field on the record page requires a user to click and applies to field generation templates, so it is neither automatic nor the right template type. Sending reps to the email composer produces an email draft and depends on manual copying, which does not meet an automated requirement. Formula fields evaluate declaratively against record data and cannot call a large language model.

**Sources:** [Create a Flex Prompt Template with Einstein AI](https://trailhead.salesforce.com/content/learn/projects/quick-start-create-prompt-builder-flex-template/build-a-flex-prompt-template), [Prompt Builder Basics](https://trailhead.salesforce.com/content/learn/modules/prompt-builder-basics)

### Question 3

Service reps at Northern Trail Outfitters must run an activated prompt template, and the security team insists that reps never see data they could not otherwise open. Which two statements should the Agentforce Specialist confirm? Choose 2 answers.

- A. Reps need Execute Prompt Templates, typically granted via the Prompt Template User permission set.
- B. Dynamic grounding uses secure data retrieval, so the prompt reflects only data the executing user is permitted to see.
- C. Templates execute in system context by default, so reps can receive summaries built from records their own profile cannot open.
- D. Reps must also hold Manage Prompt Templates, since that permission is required to execute an activated template.
- E. Field-level security can be bypassed for grounding fields by marking them as trusted prompt resources in the template.

**Correct answers:** A. Reps need Execute Prompt Templates, typically granted via the Prompt Template User permission set.; B. Dynamic grounding uses secure data retrieval, so the prompt reflects only data the executing user is permitted to see.

**Explanation:** Running a template requires the Execute Prompt Templates permission, delivered by the Prompt Template User permission set, and secure data retrieval grounds the prompt only with data the executing user is entitled to see, preserving sharing and field-level security. Prompt templates do not run in system context, which is precisely the guarantee the security team is asking about. Manage Prompt Templates is an authoring permission and is not needed to run an active template. Adding a field as a grounding resource never overrides field-level security; a user without access to the field simply does not have it merged into the prompt.

**Sources:** [Give Users Access to a Prompt Template](https://help.salesforce.com/s/articleView?language=en_US&id=ai.prompt_builder_give_users_access.htm&type=5), [Einstein Trust Layer: Designed for Trust](https://help.salesforce.com/s/articleView?id=ai.generative_ai_trust_arch.htm&language=en_US&type=5)

### Question 4

Finance at Universal Containers needs each renewal record to show a contract value calculated from line items, and the figure must match the finance system exactly every time. A stakeholder suggests generating it with a prompt template. How should the Agentforce Specialist respond?

- A. Use a prompt template and instruct the model to calculate carefully, double-check its math, and show its work.
- B. Use a prompt template grounded with a retriever over the line items so the model finds and totals the correct amounts.
- C. Use a prompt template but have a finance user manually review every generated value before it's saved to the record.
- D. Use deterministic automation, such as a roll-up or flow calculation, for the figure.

**Correct answer:** D. Use deterministic automation, such as a roll-up or flow calculation, for the figure.

**Explanation:** Prompt Builder is the right tool when the deliverable is generated language grounded in CRM data; a value that must reconcile exactly with a finance system calls for deterministic platform automation such as a roll-up summary or a flow calculation. Instructing the model to calculate carefully still leaves output that can vary between runs. A retriever improves which content the model sees but does not make arithmetic in a language model authoritative. Adding human review of every figure turns a solved automation problem into recurring manual work and still risks an incorrect number being accepted.

**Sources:** [Maximizing Generative AI with Effective Prompts](https://trailhead.salesforce.com/content/learn/modules/prompt-builder-basics/get-to-know-prompt-builder), [Prompt Builder Basics](https://trailhead.salesforce.com/content/learn/modules/prompt-builder-basics)

## Data 360 Fundamentals (20%)

### Question 5

AW Computing wants higher retrieval accuracy, and an architect proposes building an “enriched search index” as a fourth, standalone index type that exists separately from vector and hybrid indexes. What should the Agentforce Specialist correct about this plan, and what actually happens when enrichment is turned on?

- A. Enrichment is a fourth, standalone search index type that stores only large language model generated metadata, kept entirely separate from vector and hybrid indexes.
- B. Enrichment replaces the vector and hybrid options entirely, so once it is turned on for an index the underlying index type no longer matters for ranking.
- C. Enrichment is a toggle turned on when creating a vector or hybrid index, and building it then generates a plain chunk, a metadata chunk, and a question chunk for each passage.
- D. Enrichment runs only after the search index is queried, generating a fresh metadata chunk on demand for each search request instead of at index build time.

**Correct answer:** C. Enrichment is a toggle turned on when creating a vector or hybrid index, and building it then generates a plain chunk, a metadata chunk, and a question chunk for each passage.

**Explanation:** Enrichment is not a fourth index type; it is a toggle turned on when a vector or hybrid search index is created. When the index is built with enrichment on, Data 360 generates a plain chunk with the original text, a metadata chunk (keywords, entities, topic overviews, summaries), and a question chunk for each passage. Enrichment layers on top of vector or hybrid search rather than replacing it, so the underlying index type still shapes how results are scored. Enrichment happens when the index is built, not on demand, so there is no per-query metadata generation.

**Sources:** [Understanding Search Index Types in Data Cloud](https://trailhead.salesforce.com/content/learn/modules/search-index-types-data-cloud-quick-look/get-to-know-search-index-types-in-data-cloud), [Retrieve Data](https://help.salesforce.com/s/articleView?id=data.c360_a_ai_retriever_about.htm&language=en_US&type=5)

### Question 6

A new team member at Get Cloudy Consulting asks the Agentforce Specialist to define what a retriever is in Data 360. Which definition is correct?

- A. A resource that searches a search index and returns relevant results to prompt templates, agents, and flows.
- B. A data stream that continuously ingests files and Knowledge articles into Data 360 for processing.
- C. The embedding model that generates the vector representations stored inside a search index.
- D. An Einstein Trust Layer component that masks and redacts sensitive data in the prompt before it reaches the model.

**Correct answer:** A. A resource that searches a search index and returns relevant results to prompt templates, agents, and flows.

**Explanation:** A retriever is the component that searches a search index and returns relevant data to the features that ask for it, which is how prompt templates, agents, and flows receive grounding at run time. Ingestion is handled by a data stream, which is upstream of both indexing and retrieval. The embedding model produces the vectors stored in the index, but it is part of the indexing pipeline rather than the retrieval resource an admin configures. Masking sensitive data is an Einstein Trust Layer capability applied to the prompt and response, not a retrieval component.

**Sources:** [Retrieve Data](https://help.salesforce.com/s/articleView?id=data.c360_a_ai_retriever_about.htm&language=en_US&type=5)

### Question 7

Cumulus Financial needs grounded answers that return only documents for the customer's region and only material published in the last 90 days. Which two considerations should lead the Agentforce Specialist to build a search index and retriever directly in Data 360 rather than relying solely on the out-of-the-box data library configuration? Choose 2 answers.

- A. Individual retrievers support filter conditions that narrow results to a category, recency window, or region.
- B. A custom retriever lets the specialist control which fields are returned and how many results reach the prompt.
- C. Agents can only consume individual retrievers created directly in Data 360, since data libraries are designed to be called only from Flow.
- D. Content in a data library is stored outside Data 360 in a separate managed store, so data governance policies configured in Data 360 do not apply to it.
- E. A search index can only be created in an org after every existing data library in that org has first been deleted.

**Correct answers:** A. Individual retrievers support filter conditions that narrow results to a category, recency window, or region.; B. A custom retriever lets the specialist control which fields are returned and how many results reach the prompt.

**Explanation:** Search indexes are usually built for general-purpose coverage, and individual retrievers are what apply that broad index to a specific use case through filter conditions such as article category, recency, or region. A custom retriever also lets the specialist choose the fields returned and the number of results, which keeps the grounding payload relevant and within the model's context. Agents consume data libraries directly, so the claim that libraries only work in flows is false. A data library's content is stored in Data 360 and retrievers honor the org's governance policies and row-level access checks. Creating a search index has no dependency on deleting data libraries.

**Sources:** [Retrieve Data](https://help.salesforce.com/s/articleView?id=data.c360_a_ai_retriever_about.htm&language=en_US&type=5), [Upgrade AI with Real-World Data](https://trailhead.salesforce.com/content/learn/modules/agentforce-data-library-basics/upgrade-ai-with-real-world-data)

## AI Agents (35%)

### Question 8

Cloud Kicks deploys an Agentforce Service agent on an Enhanced Chat channel in an Experience site restricted to authenticated users. The agent must show a customer their own orders and nothing else. What should the Agentforce Specialist do?

- A. Ask the customer to type their email address in the chat and treat the reply as proof of identity.
- B. Rely on the large language model to filter the results so only the current customer's orders appear in the response it generates.
- C. Grant the Agent User the View All permission on Order so lookups succeed regardless of the sharing rules configured on the object.
- D. Configure the private action to confirm the session's Account and Contact IDs and an active AuthSession, then scope the order query by that contact.

**Correct answer:** D. Configure the private action to confirm the session's Account and Contact IDs and an active AuthSession, then scope the order query by that contact.

**Explanation:** On an Experience site restricted to authenticated users, each messaging session comes with an active, authenticated AuthSession and known Account and Contact IDs on the MessagingSession record, and each private action should confirm those values and that the session is still active before completing; using the session identifiers to scope the query guarantees the action returns only that customer's data. Trusting an email address typed into the chat delegates authentication to an unverified utterance. Asking the model to filter results means the records were already retrieved and a single reasoning slip exposes another customer's data. Granting View All on Order widens access for every conversation and makes over-sharing more likely, not less.

**Sources:** [Add User Identification to Agentforce Actions](https://help.salesforce.com/s/articleView?language=en_US&id=ai.service_agent_secure_actions_implement.htm&type=5)

### Question 9

Get Cloudy Consulting maintains three agents that all use the same custom action, and the team wants to change the action's instructions for only one of them. Which two facts about how topics and actions are stored should the Agentforce Specialist explain? Choose 2 answers.

- A. Adding a topic to an agent copies the topic and its actions, so each agent and agent version has its own independent copy.
- B. A custom action can be updated for one specific agent, and the update is not shared with any other agent.
- C. Editing an action on any agent immediately republishes that change to every agent that uses the action.
- D. An action can be assigned to only one topic, so each agent needs its own separately created action.
- E. Changing an action's instructions requires cloning the agent, because instructions are read-only after activation.

**Correct answers:** A. Adding a topic to an agent copies the topic and its actions, so each agent and agent version has its own independent copy.; B. A custom action can be updated for one specific agent, and the update is not shared with any other agent.

**Explanation:** When a topic is added to an agent, Salesforce makes a copy of the topic and its actions, and each agent and each agent version holds its own independent copy, which is why a change made for one agent does not leak to the others. The custom action documentation states the same thing from the action side: every agent that uses a custom action has its own copy, and updates made for one agent are not shared. The idea that an edit republishes everywhere contradicts that copy model. An action can be added to multiple topics, so there is no need to recreate it per topic. Instructions are editable on a deactivated agent, so cloning the agent is unnecessary.

**Sources:** [Add an Action to a Subagent in the Legacy Builder](https://help.salesforce.com/s/articleView?id=ai.copilot_actions_add.htm&language=en_US&type=5), [Create a Custom Agent Action in the Legacy Builder](https://help.salesforce.com/s/articleView?language=en_US&id=ai.copilot_actions_custom.htm&type=5)

### Question 10

Two builders at AW Computing disagree about whether to work in Canvas view or Script view in Agentforce Builder. What should the Agentforce Specialist tell them?

- A. Canvas view is for legacy agents only, so all new work must happen in Script view.
- B. Both views edit the same Agent Script, with Canvas summarizing it into blocks and Script view offering direct editing with syntax highlighting.
- C. Script view is mandatory for anything involving variables, because variables cannot be created, referenced, or managed without hand-written script.
- D. Switching between views creates separate agent versions that must later be reconciled.

**Correct answer:** B. Both views edit the same Agent Script, with Canvas summarizing it into blocks and Script view offering direct editing with syntax highlighting.

**Explanation:** Canvas view and Script view are two surfaces over the same Agent Script and maintain consistent agent configuration: Canvas presents the script as understandable blocks with shortcuts, such as typing a slash for common patterns and an at sign to insert resources, while Script view gives a direct editor with syntax highlighting, autocompletion, and validation. Canvas is not limited to legacy agents; the legacy builder is a separate authoring experience altogether. Variables are no longer script-only, because a low-code variable management interface is now available. Switching views does not fork the agent into competing versions, since both act on one underlying script.

**Sources:** [Agent Script | Agentforce Developer Guide](https://developer.salesforce.com/docs/ai/agentforce/guide/agent-script.html), [Learn About New Agentforce Builder with Agent Script](https://help.salesforce.com/s/articleView?id=005232662&language=en_US&type=1)

### Question 11

Cloud Kicks is rolling out a customer-facing agent on its website and on WhatsApp, with handoff to human reps when needed. Which two things should the Agentforce Specialist plan for? Choose 2 answers.

- A. Add the Service agent to enhanced Messaging channels, using Enhanced Chat for the website or app experience
- B. Expose the agent through the Agentforce panel in Lightning Experience so customers can start conversations there
- C. Route conversations, including escalation to a service rep, with an Omni-Channel flow
- D. Replace the agent with an enhanced Einstein Bot on WhatsApp, because agents support only Salesforce-hosted channels
- E. Use Bring Your Own Channel because WhatsApp is not a supported messaging channel for agents

**Correct answers:** A. Add the Service agent to enhanced Messaging channels, using Enhanced Chat for the website or app experience; C. Route conversations, including escalation to a service rep, with an Omni-Channel flow

**Explanation:** Agentforce Service agents are added to enhanced Messaging channels, with Enhanced Chat covering a company's website or mobile app, and conversations are routed with Omni-Channel flows, which is also how a complex or sensitive conversation is transferred to a live representative. The Agentforce panel in Lightning Experience is an internal employee surface, so customers never reach an agent there. WhatsApp, Facebook Messenger, LINE, Apple Messages for Business, and SMS are supported enhanced channels, so there is no need to fall back to an Einstein Bot or to Bring Your Own Channel, which exists for external messaging services that Salesforce does not provide natively.

**Sources:** [Get to Know Agentforce Service Agent](https://help.salesforce.com/s/articleView?id=service.bots_service_asa_about.htm&language=en_US&type=5)

### Question 12

A customer at Cumulus Financial requests a transfer larger than the available balance, and the business requires the agent to state the available balance, the requested amount, and the exact shortfall. Which two approaches should the Agentforce Specialist take? Choose 2 answers.

- A. Calculate the shortfall in Agent Script and pass the value into the message through a template expression.
- B. Store the balance, requested amount, and shortfall in session variables so the message is assembled from known state.
- C. Instruct the LLM to subtract the balance from the requested amount and state the result.
- D. Round the shortfall to the nearest hundred so small model arithmetic errors are not visible to the customer.
- E. Return the raw action response to the customer without any message so no numbers can be misstated.

**Correct answers:** A. Calculate the shortfall in Agent Script and pass the value into the message through a template expression.; B. Store the balance, requested amount, and shortfall in session variables so the message is assembled from known state.

**Explanation:** Exact figures must be derived in script and surfaced through template expressions that read from session variables, which leaves the model in a purely presentational role and guarantees the numbers are correct. Holding the balance, the requested amount, and the computed shortfall in session state is what makes that message assembly possible and keeps the values available for later steps. Asking the model to do the arithmetic reintroduces the risk of a wrong financial figure, which is never acceptable in a transfer scenario. Rounding to hide arithmetic errors conceals a defect instead of fixing it and misstates the amount. Dumping a raw action response gives the customer no explanation and no concrete next step.

**Sources:** [Hybrid Reasoning with New Agentforce Builder and Agent Script](https://architect.salesforce.com/docs/architect/fundamentals/guide/hybrid-reasoning-agentforce-builder-agent-script)

### Question 13

A Cloud Kicks stakeholder asks the Agentforce Specialist to explain, in one sentence, what Agent Script actually is within Agentforce Builder. Which description is accurate?

- A. A prompt template language used only inside Prompt Builder, with merge fields for grounding data and a template token limit.
- B. The language for authoring agents in Agentforce Builder, mixing natural language with expressions for conditions and actions.
- C. A replacement for Apex that lets administrators declaratively write triggers, batch jobs, and scheduled classes without deploying code.
- D. A data transformation language used to chunk, embed, and index documents before a retriever can surface them.

**Correct answer:** B. The language for authoring agents in Agentforce Builder, mixing natural language with expressions for conditions and actions.

**Explanation:** Agent Script is the authoring language for agents in Agentforce Builder, and its defining characteristic is that it mixes natural language instructions for conversational work with programmatic expressions that define if/else conditions, transitions, variables, and action selection. It is not a Prompt Builder language: prompt templates are a separate artifact that an agent action can invoke, and Prompt Builder does not use Agent Script. It does not replace Apex, because Apex, Flow, external APIs, and prompt templates all remain supported action types that Agent Script calls. Chunking and indexing belong to Data 360 and the Agentforce Data Library, not to Agent Script.

**Sources:** [Get Started with Agent Script](https://developer.salesforce.com/docs/ai/agentforce/guide/agent-script.html), [Learn About New Agentforce Builder with Agent Script](https://help.salesforce.com/s/articleView?id=005232662&language=en_US&type=1)

## Testing, Deployment, and Maintenance (10%)

### Question 14

Ursa Major Solar uploaded 200 test cases in which the expected subagent column contains values such as Order Management and Panel Warranty Lookup, matching what the builder shows on screen. Every subagent evaluation fails even though the agent visibly routes correctly. What should the Agentforce Specialist correct?

- A. Re-run the test so the nondeterministic routing settles on the expected subagent
- B. Move the routing logic into a single subagent so evaluations can match it
- C. Replace the expected subagent values with the expected response text
- D. Swap the labels for each subagent's API name, which evaluation matches

**Correct answer:** D. Swap the labels for each subagent's API name, which evaluation matches

**Explanation:** The expected subagent column is matched on the API name of the subagent, not the label displayed in the builder, so friendly names like Panel Warranty Lookup fail every comparison even when routing is right. Re-running the batch will not help, because the mismatch is a naming problem rather than nondeterministic behavior. Collapsing several subagents into one would destroy the routing design the test is meant to validate. Putting response text into the subagent column mixes up two different evaluations and still leaves subagent validation without a value it can compare.

**Sources:** [Effective Testing Setup for Agentforce Agents](https://trailhead.salesforce.com/content/learn/modules/agentforce-agent-testing/set-up-testing-criteria), [Understanding Agentforce Testing Center Results](https://trailhead.salesforce.com/content/learn/modules/agentforce-agent-testing/view-test-results)

### Question 15

Get Cloudy Consulting needs a repeatable, source-controlled path for promoting agent and prompt metadata through several orgs in a release pipeline rather than assembling change sets by hand each time. What should the Agentforce Specialist recommend?

- A. Export the agent as a CSV from Agentforce Studio and import it in each org
- B. Create the agent independently in every org and compare screenshots
- C. Copy the agent between orgs with a sandbox-to-sandbox refresh for each release
- D. Retrieve and deploy the metadata with Salesforce CLI from a version-controlled project

**Correct answer:** D. Retrieve and deploy the metadata with Salesforce CLI from a version-controlled project

**Explanation:** Salesforce CLI synchronizes metadata between an org and a local project kept in source control, which is why it is the documented alternative to click-based change sets for pipelines and CI/CD. Agentforce Studio does not offer a CSV export that reconstitutes an agent in another org. Building the agent separately in each org guarantees configuration drift and no audit trail. A sandbox refresh copies from production into a sandbox on a whole-org basis, so it cannot promote a single release forward through a pipeline.

**Sources:** [Salesforce CLI Considerations for Prompt Templates](https://help.salesforce.com/s/articleView?language=en_US&id=ai.prompt_builder_considerations_cli.htm&type=5), [Agentforce Deployment with Change Set](https://help.salesforce.com/s/articleView?id=005232697&language=en_US&type=1)

## Governance and Observability (10%)

### Question 16

At Get Cloudy Consulting, answer relevance for one subagent looks acceptable, but context relevance is consistently low. What should the Agentforce Specialist conclude?

- A. The model behind the agent is generating unsafe language, so the toxicity score under the Trust Layer view should be reviewed first.
- B. The retrieval step is returning the wrong source material, so the data sources and retriever configuration should be reviewed carefully.
- C. The agent is escalating too many conversations to a human, so the escalation rate should be reduced before anything else.
- D. The agent is ignoring the retrieved content when it writes its final response, so the instructions need to be rewritten.

**Correct answer:** B. The retrieval step is returning the wrong source material, so the data sources and retriever configuration should be reviewed carefully.

**Explanation:** Context relevance grades the accuracy of data source retrieval, so a low score with acceptable answer relevance points at what is being retrieved rather than at how the response is written, making the data sources, index, and retriever filters the place to look. Toxicity is a separate Trust Layer measure and is not implicated by relevance scores. Escalation rate is an effectiveness metric and is a symptom rather than a cause here. Responses that ignore retrieved content would instead depress answer faithfulness, which is not the dimension flagged.

**Sources:** [Monitor Agent Performance Metrics and Quality Scores](https://trailhead.salesforce.com/content/learn/modules/agent-analytics-and-monitoring/monitor-agent-metrics-and-scores), [Investigate Agent Sessions and Intents](https://trailhead.salesforce.com/content/learn/modules/agent-analytics-and-monitoring/investigate-agent-sessions-and-intents)

### Question 17

A steering committee at Universal Containers asks how the team will judge whether its new agent is succeeding once it is live. What should the Agentforce Specialist recommend?

- A. Compare production behavior against the pass rate recorded by Testing Center before deployment, treating that number as the agent's ongoing success target.
- B. Judge success by the number of sessions handled, because volume is the clearest and easiest signal of adoption to report to leadership.
- C. Agree on KPI-driven goals during the Ideate phase, then track them against Agent Analytics after deployment and iterate.
- D. Wait for customer feedback ratings to accumulate for a full quarter before drawing any conclusion, so the sample is large enough to trust.

**Correct answer:** C. Agree on KPI-driven goals during the Ideate phase, then track them against Agent Analytics after deployment and iterate.

**Explanation:** Monitoring is most effective when the KPI-driven goals are set during Ideate so that production metrics can be measured against an agreed target and drive the next iteration. A Testing Center pass rate describes behavior on authored test cases and cannot stand in for live effectiveness or quality. Session volume is a usage metric and says nothing about whether conversations were resolved or escalated. Deferring judgment for a quarter ignores the effectiveness, health, and quality metrics available continuously, and it delays the optimization loop the lifecycle is built around.

**Sources:** [Get Started with Agent Monitoring](https://trailhead.salesforce.com/content/learn/modules/agent-analytics-and-monitoring/get-started-with-agent-monitoring), [Monitor Agent Performance Metrics and Quality Scores](https://trailhead.salesforce.com/content/learn/modules/agent-analytics-and-monitoring/monitor-agent-metrics-and-scores)

## Multi-Agent Orchestration (5%)

### Question 18

Get Cloudy Consulting proposes standing up five separate expert agents purely to add a variety of perspectives to one generative insight. Before approving the design, the Agentforce Specialist wants to evaluate the lighter-weight alternatives Salesforce guidance says to consider first. Which two approaches should be evaluated? Choose 2 answers.

- A. Deploy each expert agent in its own Salesforce org so the perspectives can be coordinated with the agent-to-agent protocol.
- B. Specify multiple prompts and let the system work as an ensemble of prompts instead of defining multiple expert agents.
- C. Replace the generative insight with a static report so no reasoning is required at all.
- D. Ground the response through actions that access contextually appropriate data.
- E. Configure one agent per model vendor so that each perspective comes from a different provider by design.

**Correct answers:** B. Specify multiple prompts and let the system work as an ensemble of prompts instead of defining multiple expert agents.; D. Ground the response through actions that access contextually appropriate data.

**Explanation:** Salesforce's guidance on the Model of Models pattern explicitly tells architects to consider lighter-weight approaches before defining several expert agents: use an ensemble of prompts rather than multiple agents, and lean on grounding through actions that retrieve contextually appropriate data. Spreading the experts across separate orgs multiplies the cross-org trust, identity, and observability work without addressing whether multiple agents are warranted at all. Substituting a static report abandons the generative requirement instead of simplifying how it is met. Creating an agent per model vendor is a characteristic of the heavyweight patterns being questioned, not a lighter alternative to them.

**Sources:** [Enterprise Agentic Architecture and Design Patterns](https://architect.salesforce.com/docs/architect/fundamentals/guide/enterprise-agentic-architecture.html)

## Related pages

- [Salesforce Agentforce Specialist practice exam](https://www.certifyforce.com/salesforce-agentforce-specialist-practice-exam)
- [Practice exam packs](https://www.certifyforce.com/practice-exams)
- [Official Agentforce Specialist exam guide](https://help.salesforce.com/s/articleView?id=005298924&language=en_US&type=1)