Salesforce Certified Agentforce Specialist – Complete 2026 Guide

Correct answer: B. Managing and optimizing Agentforce with deep platform and Agentforce knowledge

Certified Agentforce Specialists manage and optimize Agentforce and have deep understanding of both Salesforce platform configuration and Agentforce capabilities.

• Writing Apex code is a developer responsibility, not the core focus of the Agentforce Specialist role, which is configuration-focused.
• Marketing campaign design is a Marketing Cloud function, unrelated to Agentforce agent configuration.
• CPQ product configuration is a separate specialization, not part of the Agentforce Specialist's core responsibility.

Correct answer: B. A defined area of expertise that tells an agent when and how to handle a specific type of customer request, along with the actions available for that topic

Topics define the agent's scope, each topic specifies the instructions, tone, and available actions for a particular type of request such as order management or billing enquiries.

• Topics aren't related to report/dashboard grouping, that's an analytics concept, not an agent configuration concept.
• A Flow trigger condition is a distinct automation concept unrelated to how Topics scope an agent's conversational behavior.
• Permission set groups control data/feature access, not which types of customer requests an agent recognizes and handles.

Correct answer: B. A discrete task the agent can perform, such as running a Flow, calling an Apex method, invoking an external API, or generating content from a Prompt Template

Agent Actions are the executable steps available within a topic. Supported action types include Flows, Apex, external services, prompt templates, and standard Salesforce actions.

• Page layout configuration is a UI/admin concept unrelated to what an agent can actually do during a conversation.
• An automated email on case escalation is a standard Salesforce automation feature, not an Agentforce-specific Agent Action.
• A custom LWC in the chat window is a UI display component, not a task the agent performs.

Correct answer: B. Data Cloud

Data Cloud provides unified customer profiles that Agentforce can query to ground agent responses with accurate, real-time customer context.

• Marketing Cloud Journeys manages marketing automation sequences, not unified customer data context for agent grounding.
• Commerce Cloud's product catalog is scoped to product data, not the broader unified customer profile Agentforce needs.
• Slack channels are a communication surface, not a data source for unified customer context.

Correct answer: B. A security framework that masks PII before sending prompts to the LLM, enforces zero data retention, and provides audit logs of all AI interactions

The Einstein Trust Layer ensures that customer data is protected when Agentforce interacts with external LLMs, masking sensitive fields, preventing data retention by model providers, and logging all AI activity for audit.

• A KPI dashboard reports on performance after the fact, it doesn't provide the real-time privacy and security protections the Trust Layer delivers.
• Requiring human sign-off on every response is a workflow/approval process, not what the Einstein Trust Layer itself does.
• Custom Flow-based output validation is a configuration choice an admin might build, it's not what the platform's built-in Trust Layer provides.

Correct answer: B. Instructions that define what the agent should never do, for example, never discuss competitor pricing, never make promises about refunds

Guardrails are negative instructions that constrain agent behaviour, preventing it from going off-topic, making unauthorised commitments, or generating harmful content.

• Row-level security controls record visibility generally; Guardrails are specifically about defining prohibited agent behaviors, not data access.
• Message rate limits are a technical throttling mechanism, unrelated to defining what topics or promises an agent should avoid.
• A/B testing routes traffic between agent versions for experimentation, it doesn't define behavioral constraints.

Correct answer: B. Experience Cloud sites, messaging channels (SMS, WhatsApp), and web chat widgets

Agentforce agents can be deployed across multiple channels including Experience Cloud, web chat, and messaging platforms, enabling self-service without needing to reach a human agent.

• Restricting agents to only the internal Service Console would defeat the purpose of self-service customer-facing deployment.
• Limiting deployment to only Slack and email excludes the web chat and Experience Cloud channels Agentforce actually supports.
• Restricting to only mobile push notifications ignores the broader self-service channel options actually available.

Correct answer: B. Creating and managing reusable prompt templates that merge static instructions with dynamic CRM and Data Cloud data for agent actions

Prompt Builder enables admins and specialists to create prompt templates that combine instructions with live CRM data, used for field generation, record summaries, sales emails, and agent actions.

• Lightning Record Page layout building is a separate Salesforce admin task, unrelated to Prompt Builder's function.
• Omni-Channel routing configuration is a Service Cloud feature, distinct from Prompt Builder's prompt-template role.
• Writing Apex test classes is a developer task unrelated to Prompt Builder, which is a declarative prompt-authoring tool.

Correct answer: B. It uses a large language model to classify the customer's intent and match it to the most relevant Topic based on the topic's description and scope

Agentforce's reasoning engine (the LLM) classifies user intent and selects the most relevant Topic based on the topic descriptions and scope configured by the specialist.

• Case category field matching is a simpler rule-based approach, not how Agentforce's LLM-driven Topic classification actually works.
• Requiring the customer to manually pick a topic from a dropdown removes the natural-language understanding that defines Agentforce's approach.
• Random topic assignment with a fixed escalation threshold isn't how Topic matching works, it's driven by LLM intent classification, not chance.

Correct answer: B. Review and refine the Topic instructions, tighten guardrails, improve prompt template grounding, and test with diverse conversation scenarios

Tuning agent performance requires iterative refinement of topic instructions, guardrails, and prompt templates, alongside testing with real-world conversation variations to identify and fix gaps.

• Deleting and rebuilding from scratch discards existing configuration work without first diagnosing what's actually misconfigured.
• Adding more Agent Actions gives the agent more options but doesn't fix the underlying issue of incorrect or off-topic responses.
• Increasing LLM temperature adds randomness to output, which would make incorrect or off-topic responses more likely, not less.

Correct answer: B. The Escalate to Agent standard action, which transfers the conversation and full context to Omni-Channel routing

Agentforce includes a standard escalation action that hands the conversation, including full transcript and context, to a human agent via Omni-Channel, so the handover is not restarted.

• 'Agent Escalation Flow' isn't the actual named Agentforce capability, the real standard action is 'Escalate to Agent.'
• A Workflow Rule triggering Case creation on negative sentiment is a possible automation pattern, but it isn't the built-in mechanism for live conversation hand-off.
• An outbound email notification doesn't transfer the live conversation and context, it just alerts a team after the fact.

Correct answer: B. Configuring and optimising Agentforce agents, including topics, actions, prompt templates, guardrails, channels, and Data Cloud grounding

The Agentforce Specialist exam tests the ability to declaratively configure agents: defining topics and actions, crafting prompt templates, applying guardrails, deploying to channels, and grounding agents with Data Cloud.

• Writing Apex and LWC for custom integrations is developer work, not the Specialist certification's configuration focus.
• Data Cloud ingestion pipeline architecture is a Data Cloud/architect-level skill, not the core Agentforce Specialist exam focus.
• Administering Omni-Channel routing and Case Classification are separate Service Cloud admin skills, not Agentforce-specific configuration.

Correct answer: B. To define negative constraints, what the agent must never do

Guardrails are negative constraints that prevent the agent from taking certain actions or making certain statements.

• Guardrails don't affect response creativity, that's unrelated to their purpose.
• Guardrails are configuration settings, not an LLM temperature control.
• Guardrails define behavioral limits; they have nothing to do with adding data sources.

Correct answer: B. Flow action; agents can invoke Flows as discrete actions

Flow is one of the standard action types an agent can use to perform automated tasks.

• Agents can invoke more than just Apex; Flow is itself a distinct, supported action type.
• External APIs are a separate action type from Flow; 'only external APIs' incorrectly excludes Flow actions.
• Prompt templates are a different action type from Flow invocation, they don't cover invoking a Flow directly.

Correct answer: B. When the agent needs access to unified customer profiles and real CRM data for grounded responses

Data Cloud grounding provides the agent with unified customer data for accurate, data-backed responses.

• Grounding is a core, commonly used Agentforce capability; 'never' misrepresents its actual purpose and use.
• Data Cloud grounding isn't restricted to external channels only, it applies whenever the agent needs real customer data, regardless of channel.
• Grounding isn't a testing-only feature, it's meant for production use whenever real CRM context improves response accuracy.