Salesforce Certified Agentforce Specialist (AI-201) Exam Topics and Questions
These Salesforce Certified Agentforce Specialist (AI-201) exam topics are organized according to official exam domains to help candidates quickly verify coverage and focus on assessment rather than theory. Each domain is paired with topic-wise Salesforce Certified Agentforce Specialist (AI-201) sample questions that reflect how objectives are tested in the actual exam. This structure enables efficient review, targeted self-assessment, and rapid identification of weak areas when preparing for the Salesforce Certified Agentforce Specialist certification exam.
Let's Practice Free Salesforce Certified Agentforce Specialist (AI-201) Questions Aligned with Official Exam Topics
Prompt engineering sits at the front of the exam because it controls whether your AI outputs help or confuse. You will need to know when Prompt Builder solves a problem and when it does not, how access and security shape what a prompt can touch, and how grounding keeps a model honest. The cost comes in picking the right template type and writing instructions that produce consistent results under production load. β Template selection and access Prompt Builder is not...
Prompt engineering sits at the front of the exam because it controls whether your AI outputs help or confuse. You will need to know when Prompt Builder solves a problem and when it does not, how access and security shape what a prompt can touch, and how grounding keeps a model honest. The cost comes in picking the right template type and writing instructions that produce consistent results under production load.
β Template selection and access
Prompt Builder is not a universal tool. You need to recognise the scenarios where field generation or flex templates fit, and where they create more complexity than they resolve. Each template type carries constraints on what it can return and how it integrates with the rest of the platform. Access controls govern who can build, edit, activate, and execute a template, and those permissions flow through profiles, permission sets, and sharing rules in ways that are easy to misconfigure. A prompt that works in your sandbox may fail in production because the running user lacks visibility on the data the grounding pulls. The exam expects you to map business requirements to the right template and to anticipate access failures before they reach end users.
β Grounding, lifecycle, and technique
Grounding techniques determine where the model draws its facts. You will see scenarios that need retrieval from Data Cloud, from standard objects, or from external sources, and you must choose the method that balances accuracy with latency. Creating a template is straightforward; activating and executing it correctly under varying user contexts is not. The lifecycle from draft to active to retired matters because versioning and change tracking are light, and a bad activation can propagate across records faster than you can roll it back. Writing effective prompts means being specific about role, task, format, and constraints, then testing whether the output degrades when input variety increases. The Trust Layer enforces data masking, toxicity filtering, and audit logging, but you still need to know which controls apply at inference time and which apply at template design time. Model management includes blocking access to certain models entirely, and the exam will ask when and how to enforce that restriction.
How Prompt Engineering is tested
Twenty percent of the exam means roughly one question in five. Items present a business requirement or a technical constraint and ask which template type, grounding approach, or security control applies. You lose marks by confusing field generation with flex, by recommending grounding that cannot see the necessary data, or by missing an access control gap that would block execution. Scenario questions describe a prompt that fails in production and expect you to identify whether the issue is permissions, grounding configuration, model access policy, or prompt construction. The demand is not to write a perfect prompt from scratch but to diagnose why one does not behave as intended and to select the fix that aligns with Trust Layer requirements. Candidates who treat all templates as interchangeable, or who assume grounding always works if the data exists somewhere in the org, lose marks quickly.
The practice test includes prompt engineering scenarios across every grounding technique and template type, so you see the decision patterns before the real exam. The PDF gives you the same question set for offline review when you need to check coverage without opening a browser.
The scenario below turns on choosing the correct grounding method when data visibility and latency both constrain the solution.
Universal Containers implements Custom Agent Actions to enhance its customer service operations. The development team needs to understand the core components of a Custom Agent Action to ensure proper configuration and functionality. What should the development team review in the Custom Agent Action configuration to identify one of the core components of a Custom Agent Action?
Data 360 underpins retrieval, and retrieval underpins every agent that needs to answer from organisational knowledge rather than invention. The exam expects you to understand how the Agentforce Data Library organises content, how chunking and indexing prepare that content for search, and how retrievers match a query to the right fragment. Mistakes here propagate into every downstream topic, because an agent with poor retrieval hallucinates or returns irrelevant answers no matter how well the rest of the architecture is built....
Data 360 underpins retrieval, and retrieval underpins every agent that needs to answer from organisational knowledge rather than invention. The exam expects you to understand how the Agentforce Data Library organises content, how chunking and indexing prepare that content for search, and how retrievers match a query to the right fragment. Mistakes here propagate into every downstream topic, because an agent with poor retrieval hallucinates or returns irrelevant answers no matter how well the rest of the architecture is built.
How Data 360 Fundamentals is tested
Another twenty percent weight, so expect similar volume to the first topic. Questions describe a retrieval failure or a performance problem and ask which part of the Data 360 pipeline is misconfigured. You will see scenarios where chunking is too coarse and context is lost, or too fine and search becomes noisy. Indexing delays, stale content, and retriever selection all appear as plausible distractors. The exam also tests whether you know what belongs in the Data Library and what does not: not every Salesforce object or external file is a good candidate, and adding everything degrades quality. Candidates who assume indexing is automatic, or who think chunking size does not matter, lose marks. The demand is to trace a retrieval outcome back to its root cause in chunking, indexing, or retriever logic, and to propose a fix that does not break something else.
Working through retrieval questions in timed conditions shows you whether you can distinguish chunking problems from indexing problems under pressure. The demo lets you try a subset before committing, so you can confirm the coverage matches what you need to fill.
The question that follows asks you to identify the cause of poor retrieval quality when the Data Library configuration is incomplete.
Which object stores the conversation transcript between the customer and the agent?
AI Agents carries more than a third of the total weight, and the breadth matches the percentage. You need to know how agent script orchestrates reasoning, how hybrid reasoning balances generative and deterministic paths, and how to configure topics and actions so the agent solves the right problem without veering into adjacent ones. Channel integration, security context, agent type selection, and API use all appear here. The cost is in the detail: a misconfigured filter or a missing variable can...
AI Agents carries more than a third of the total weight, and the breadth matches the percentage. You need to know how agent script orchestrates reasoning, how hybrid reasoning balances generative and deterministic paths, and how to configure topics and actions so the agent solves the right problem without veering into adjacent ones. Channel integration, security context, agent type selection, and API use all appear here. The cost is in the detail: a misconfigured filter or a missing variable can send the agent down the wrong branch, and a security context mismatch can block an action that should succeed.
β Agent architecture and reasoning
Agent script defines the flow: which topics are in scope, which actions are available, and how the agent decides what to do next. The building blocks are topics, actions, and the reasoning layer that connects them. Hybrid reasoning combines generative decision-making with deterministic rules, and you control the balance through filters, variables, and template expressions. Canvas view gives you a visual representation; Script view gives you the underlying logic. Both are examinable, and you need to move between them without losing track of what the agent will actually do at runtime. Deterministic behaviour is managed by constraining topic scope, setting variable defaults, and writing template expressions that produce predictable outputs. The exam expects you to recognise when a scenario needs tighter control and which mechanism delivers it without overcomplicating the script.
β Topics and actions
Standard topics cover common patterns like case deflection and knowledge search. Custom topics extend the agent into your specific processes. Standard actions include record lookups, updates, and flows; custom actions wrap Apex, external services, or anything else you can invoke. Selecting the right combination means understanding what each topic and action can do, what permissions it requires, and what happens when it fails. Configuration includes setting input and output mappings, defining fallback behaviour, and deciding whether the agent should surface the action to the user or execute it silently. A misconfigured action mapping is a common trap, and the exam will show you scenarios where the agent has access to the action but cannot use it because the inputs are not populated or the security context is wrong.
β Channels, context, and agent type
Agents connect to digital experience sites, email, voice, and Slack, and each channel imposes different constraints on what the agent can display and how the user interacts. Voice limits rich formatting; email introduces latency; Slack requires app configuration and token management. The security context in which the agent runs determines which records it can see and which actions it can execute. That context may be the running user, a named service account, or a guest user profile, and the choice affects everything downstream. Employee agents and Service agents differ in scope, permissions, and typical use cases. Knowing when to use each means matching the agent type to the audience and the data it needs to touch. Agent API is the programmatic route for embedding agent behaviour into custom applications, and scenarios will test whether you recognise when API invocation is the right pattern instead of a channel-based deployment.
How AI Agents is tested
Thirty-five percent means this topic dominates the item count. Expect scenario-based questions that describe an agent misbehaving and ask which part of the configuration is wrong. You will see cases where the agent triggers the wrong topic because filters are too loose, where an action fails because the security context lacks access, and where a channel limitation breaks the user experience. The exam tests your ability to map a use case to the correct agent type, to configure hybrid reasoning so generative and deterministic paths work together, and to debug a deployment where the agent works in sandbox but fails in production because permissions or data visibility changed. Candidates lose marks by choosing custom actions when standard ones suffice, by ignoring security context, and by assuming all channels behave identically. The demand is to hold the entire agent lifecycle in your head and to identify the single misconfiguration that explains the failure.
Volume matters here because agent scenarios branch in many directions, and you need enough repetitions to spot the pattern behind topic selection, action mapping, and security context errors. The practice test gives you that volume under exam conditions, and the free demo covers a sample of agent questions so you can verify the difficulty before you buy.
The scenario that follows requires you to choose the correct agent type and channel configuration when both audience and platform constraints apply.
Testing Center is where you validate agent behaviour before it reaches production. The exam expects you to know how to construct test cases, how evaluations score agent responses, and what the results tell you about topics, actions, and reasoning. Deployment from sandbox to production is not a metadata push and forget: agents and templates both carry dependencies that break in subtle ways if you miss a step. Ten per cent weight means fewer questions, but each one can cost you...
Testing Center is where you validate agent behaviour before it reaches production. The exam expects you to know how to construct test cases, how evaluations score agent responses, and what the results tell you about topics, actions, and reasoning. Deployment from sandbox to production is not a metadata push and forget: agents and templates both carry dependencies that break in subtle ways if you miss a step. Ten percent weight means fewer questions, but each one can cost you if you have not deployed an agent under real constraints.
β Testing and evaluation
Testing Center lets you simulate conversations and measure whether the agent selects the correct topic, invokes the right action, and produces an acceptable answer. You build test cases by defining user inputs and expected outcomes, then run evaluations to score the agent's performance. Evaluations use metrics like topic accuracy, action success rate, and response quality, and the exam expects you to interpret those metrics and trace a failure back to a configuration problem. A low topic accuracy score means the agent is misclassifying user intent; a low action success rate points to permissions, input mappings, or missing data. Testing is not a one-time gate before go-live: you run it after every script change, every new topic, and every action addition, because agent behaviour is emergent and small changes propagate unpredictably.
β Deployment considerations
Deploying an agent from sandbox to production requires moving the agent definition, all referenced topics and actions, any custom prompts, the Data Library configuration, and the security settings that govern execution context. Miss one dependency and the agent fails silently or behaves differently in production. Templates have their own deployment path, and you need to check that grounding sources, model access policies, and Trust Layer settings match across environments. The exam tests whether you know which components move via change sets, which require manual reconfiguration, and which need post-deployment validation. Candidates who assume agent deployment is a single metadata retrieve and deploy operation lose marks when they miss the manual steps or the environment-specific settings that do not transfer.
How Testing, Deployment, and Maintenance is tested
Ten percent means roughly one question in ten. Scenarios describe an agent that passed all sandbox tests but fails in production, and you must identify what was missed during deployment. You will see questions on how to interpret Testing Center evaluation results, how to construct a test case that isolates a specific topic or action, and what to check when an agent behaves differently across environments. The exam also covers ongoing maintenance: when to re-run evaluations, how to handle version drift between sandbox and production, and what happens when a dependency like a flow or an Apex class changes underneath a deployed agent. Candidates lose marks by assuming tests are optional, by deploying without validating dependencies, and by ignoring evaluation metrics that signal a problem. The demand is to treat testing and deployment as continuous disciplines, not one-time gates.
Deployment questions often turn on a single overlooked dependency or a misread evaluation metric. Practising those scenarios helps you build the checklist you will need when you deploy your first agent after certification, and the PDF format means you can review the steps offline without needing a Salesforce org open.
The question below asks you to identify the missing step when an agent deploys successfully but fails to execute actions in the target environment.
Governance and observability keep agents running well after go-live. You need to know how to monitor agent performance, how to interpret analytics that show where users drop off or where the agent fails, and how to optimise based on those signals. Ten per cent weight and only two sub-topics mean the exam expects depth on monitoring and analytics rather than breadth across many tools. How Governance and Observability is tested Managing and monitoring agents means tracking conversation volume, topic hit...
Governance and observability keep agents running well after go-live. You need to know how to monitor agent performance, how to interpret analytics that show where users drop off or where the agent fails, and how to optimise based on those signals. Ten percent weight and only two sub-topics mean the exam expects depth on monitoring and analytics rather than breadth across many tools.
How Governance and Observability is tested
Managing and monitoring agents means tracking conversation volume, topic hit rates, action success rates, and user satisfaction scores. The exam expects you to know which metrics indicate a healthy agent and which signal a problem that needs intervention. Agent analytics surface patterns: topics that are invoked but never complete, actions that fail consistently, and user queries that the agent cannot classify. Optimisation is the response to those patterns, and it may mean refining topic scope, rewriting prompts, adjusting grounding, or adding new actions. Questions describe an analytics report or a monitoring dashboard and ask what the data tells you and what you should do next. Candidates lose marks by choosing optimisation steps that do not address the root cause, by ignoring metrics that clearly show a failure mode, and by assuming that agent performance is static once deployed. The demand is to read the signals, diagnose the issue, and select the intervention that improves outcomes without introducing new problems.
Analytics and optimisation questions require you to interpret data and choose the right fix. Seeing those scenarios in the practice test builds the pattern recognition you need to move quickly through them on exam day, and working under timed conditions shows you whether you can hold the full context in your head while the clock runs.
The scenario that follows presents agent analytics showing a drop in action success rates, and you must identify the most likely cause and the correct remediation step.
Multi-agent orchestration is the smallest topic by weight, but it tests whether you recognise when a single agent is not enough and when splitting responsibility across multiple agents improves scalability, control, or specialisation. You also need to know the purpose of open standard protocols like MCP and A2, which govern how agents communicate and hand off tasks. Five per cent means only a handful of questions, but they expect you to distinguish architectural patterns that look similar at first glance....
Multi-agent orchestration is the smallest topic by weight, but it tests whether you recognise when a single agent is not enough and when splitting responsibility across multiple agents improves scalability, control, or specialisation. You also need to know the purpose of open standard protocols like MCP and A2, which govern how agents communicate and hand off tasks. Five percent means only a handful of questions, but they expect you to distinguish architectural patterns that look similar at first glance.
How Multi-Agent Orchestration is tested
Scenarios describe a use case and ask whether a multi-agent architecture is justified or whether a single agent with more topics would suffice. The exam expects you to weigh the trade-offs: multiple agents increase complexity and require orchestration logic, but they allow finer-grained security contexts, independent scaling, and clearer separation of concerns. You lose marks by recommending multi-agent when the problem does not justify the overhead, or by sticking to a single agent when the security or scaling requirements make that unworkable. Questions on MCP and A2 test whether you understand what these protocols enable and when they matter. The demand is not to implement orchestration from scratch but to recognise the architectural inflection point where one agent becomes many, and to know which protocol supports the handoff pattern you need.
Architectural questions reward pattern recognition, and you build that by seeing enough scenarios to distinguish when multi-agent is the right call and when it is over-engineering. The question bank includes examples of both, and the free demo gives you a sample so you can check the reasoning style before you commit.
The question below describes a use case with conflicting security requirements and asks whether a multi-agent architecture is appropriate and which protocol applies.
Ready to Start Practicing?
Access all questions and start your exam preparation journey
Upgrade to Full Salesforce Certified Agentforce Specialist (AI-201) Exam Questions π