Overview
AI Agents carries 35% of the exam and eight objectives, making it the heaviest section by far. It covers how an Agentforce agent works end to end. The reasoning engine (Atlas) classifies a user's request into a topic, follows that topic's instructions and chooses actions such as flows, Apex, prompt templates or standard actions to complete the job.
The Spring '26 guide adds Agent Script, the language behind the new Agentforce Builder, which you can edit in Canvas view or Script view. Agent Script makes hybrid reasoning possible: deterministic steps such as required checks and fixed sequences run exactly as written, while the model handles open conversation. Related objectives cover filters, variables and template expressions for control, choosing standard and custom topics and actions, deploying agents to channels such as Experience Cloud sites, email, voice and Slack, the security context of the agent user, choosing Employee or Service agents, and when to use the Agent API.
How it's tested
Most questions describe an agent that misbehaves or a business process to automate. You decide which layer to change: topic classification description, instructions, an action, a variable with a filter, a scripted deterministic step, the agent user's permissions, the channel setup, or the agent type itself.
-
Right layer to fix
An agent picks the wrong topic, skips a step or calls the wrong action, and you choose whether to edit descriptions, instructions, filters or script logic.
-
Agent type and channel
Users and channels are described, and you choose Employee or Service agent and the deployment path such as messaging, email, voice or Slack.
-
Permissions and integration
An action fails or exposes too much data, or an external app needs the agent, and you reason about the agent user or Agent API.
A bank's service agent must verify a customer's identity before discussing balances, but in testing it sometimes skips verification when a customer sounds urgent. Instructions alone are probabilistic. Setting a Verified variable when the verification action succeeds, and adding a filter so the balance action is available only when Verified is true, makes the order deterministic. In Agent Script the same rule can be written as a required step.
Objectives 8
Official objectives from the Salesforce outline, in outline order.
3.1 Explain how an agent works and its basic building blocks of agent script.Topics, instructions, actions and the reasoning loop.
- What it means
- An Agentforce agent receives an utterance, the reasoning engine classifies it into a topic using topic names and classification descriptions, then follows that topic's scope and instructions to choose actions. Results come back to the engine, which may call more actions or respond. Agent Script expresses these building blocks, including topics, instructions, actions, variables and logic, in one readable definition.
- What to understand
- Understand the loop: classify, plan, act, observe, respond. Topics group related jobs; instructions guide behavior within a topic; actions do the work, and their names, descriptions and input and output descriptions are what the model reads when choosing them. Agent Script brings these elements together so builders can see and edit the full agent definition, including deterministic logic, rather than scattered settings.
- How it shows up
- An agent chooses the wrong topic or action, and you identify which building block's description or configuration needs changing.
- Common confusion
- Thinking the model reads action code to decide; it relies mainly on names and descriptions.
3.2 Explain the components and benefits of hybrid reasoning, including how Agent Script functions in Canvas and Script View.Combining scripted logic with model reasoning.
- What it means
- Hybrid reasoning lets an agent mix deterministic steps written in Agent Script with flexible model-driven reasoning. In the Agentforce Builder, Canvas view presents the agent's logic visually and in plain language, while Script view shows the underlying Agent Script code. Both edit the same definition.
- What to understand
- The benefit is predictability where it matters and flexibility elsewhere. Business rules such as required order of steps, conditional branches and handoffs can be scripted so they run the same way every time, while open questions remain with the reasoning engine. Builders who prefer low-code work in Canvas view; developers can review and change code in Script view, and the definition can be kept in source control. Choose scripted logic when a step must never be skipped.
- How it shows up
- A process requires strict sequencing within a conversation, and you choose a scripted step over more instructions.
- Common confusion
- Treating Canvas view and Script view as separate agents or versions rather than two views of one definition.
3.3 Given a use case, manage deterministic behavior for the agent using mechanisms like filters, variables, and template expressions.Variables, conditional filters and expressions for control.
- What it means
- Variables hold values during a session, either set by actions or mapped from context such as the messaging session or user record. Filters use variables to decide whether a topic or action is available. Template expressions insert variable values into instructions or scripted logic so behavior adapts to known facts.
- What to understand
- Use these mechanisms when a rule must hold regardless of how the model interprets the conversation. Typical cases are gating sensitive actions until verification succeeds, limiting a topic to certain customer tiers, or passing a known record ID to an action instead of asking the user. Context variables come from the channel or user, while custom variables are set by action outputs. Filters remove options from the model's choice, which is stronger than telling it not to use them.
- How it shows up
- A sensitive action must run only after a condition is met, and you choose variables and filters instead of instructions.
- Common confusion
- Relying on instruction wording such as always verify first, which the model may not follow every time.
3.4 Given a scenario, select and configure standard topics, custom topics, standard Agent actions, and custom Agent actions.Reusing standard building blocks and building custom ones.
- What it means
- Agentforce includes standard topics and actions from the Agentforce Asset Library, such as general CRM, case management, knowledge answers, and record query and update actions. Custom topics define new jobs with their own scope and instructions, and custom actions are built from flows, Apex invocable methods, prompt templates or external services.
- What to understand
- Start from standard topics and actions when they match the job, then extend. Custom actions need clear names, descriptions and input and output instructions, and should be marked to require user confirmation or show results when appropriate. Topics should be distinct, with classification descriptions that do not overlap, and contain only the actions relevant to that job. Too many topics or overlapping descriptions cause misrouting.
- How it shows up
- A business process is described, and you decide whether a standard topic or action fits or which custom action type to build.
- Common confusion
- Building a custom Apex action for a record lookup that a standard action or flow already provides.
3.5 Explain the process for connecting agents to various channels such as digital experience, email, voice, and Slack.Deploying agents where users work.
- What it means
- Service agents usually reach customers through Messaging for In-App and Web on websites and Experience Cloud sites, routed by Omni-Channel flows, and can also handle email and voice through Agentforce Voice. Employee agents appear in Lightning Experience, the Salesforce mobile app and Slack.
- What to understand
- Know the main steps for each channel. For messaging, create a messaging channel and an Omni-Channel flow that routes to the agent, then deploy the embedded service on the site. Email needs an email channel configured for the agent. Voice requires Agentforce Voice configured with contact center settings. Slack deployment connects an employee agent to a Slack workspace. Add escalation to human agents through Omni-Channel where customers may need a person.
- How it shows up
- A company wants its agent on a website, in email or in Slack, and you choose the configuration step that is missing.
- Common confusion
- Expecting an agent to appear on a site just by activating it, without a messaging channel and routing.
3.6 Explain the security context in which the agent is actually running, and how it impacts agent action execution.Agent user versus running user permissions.
- What it means
- Service agents run as a dedicated agent user that has its own profile and permission sets; customers talking to the agent are not Salesforce users. Employee agents run in the context of the signed-in employee, so they can do only what that employee can do.
- What to understand
- Every action respects the permissions of the user it runs as. For Service agents, the agent user must have object and field access for each action and nothing more, following least privilege. Flows set to run in system context and Apex without sharing can bypass those checks, so review them before adding them as actions. For Employee agents, actions fail or return nothing when the employee lacks access, which is usually correct. Use variables and filters to scope data to the verified customer.
- How it shows up
- An action works in testing but fails in production, or exposes too much data, and you trace it to the running user.
- Common confusion
- Assuming a Service agent runs as the customer or as an admin, rather than as its dedicated agent user.
3.7 Given a scenario, identify when to use an Employee or Service agent.Choosing the agent type by audience and channel.
- What it means
- Employee agents help internal users, such as sellers or service reps, inside Salesforce and Slack using their own access. Service agents talk to external customers on digital channels, voice and email, running as an agent user with escalation to humans.
- What to understand
- Decide by audience first. Internal users who need help with records, summaries or tasks get an Employee agent. External customers who need answers or self-service get a Service agent. Consider authentication, channel and data scope: customer-facing agents need verification and narrow permissions, while employee agents inherit the employee's sharing.
- How it shows up
- A requirement describes users and channels, and you choose the agent type and justify it.
- Common confusion
- Picking a Service agent for internal staff because the use case involves cases or service records.
3.8 Given a scenario, identify when it's appropriate to use Agent API.Calling agents from external applications.
- What it means
- The Agent API is a REST API that lets an external application start a session with an Agentforce agent, send messages and receive responses. It authenticates through an external client app or connected app using the client credentials flow.
- What to understand
- Use the Agent API when a custom front end, mobile app, backend system or another platform must talk to an agent and no standard channel fits. It is not needed when Messaging for In-App and Web, Slack or Lightning already provides the experience. Remember that the API calls an agent; it does not replace topics, actions or permissions, which still apply.
- How it shows up
- A company has its own app or system that needs agent responses, and you decide whether the Agent API or a standard channel fits.
- Common confusion
- Choosing the Agent API for a website chat that the standard embedded messaging deployment already supports.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Atlas Reasoning Engine
The engine that classifies requests, plans and selects actions.
- Remember
- It reasons over topic and action descriptions.
- Tested as
- How an agent works questions.
- Trap
- Assuming it reads action code.
-
Topic
A job the agent handles, with classification description, scope and instructions.
- Remember
- Distinct, non-overlapping descriptions drive correct routing.
- Tested as
- Misrouting scenarios.
- Trap
- Creating many overlapping topics.
-
Instructions
Natural language guidance inside a topic.
- Remember
- Guides behavior but can be overridden by the model.
- Tested as
- Behavior tuning questions.
- Trap
- Using instructions for rules that must always hold.
-
Agent action
Work the agent performs, built from flows, Apex, prompt templates or standard actions.
- Remember
- Names and descriptions guide action selection.
- Tested as
- Configuration questions.
- Trap
- Vague action descriptions.
-
Agent Script
Language that defines an agent's topics, actions, variables and deterministic logic.
- Remember
- Editable in Canvas view or Script view of the Agentforce Builder.
- Tested as
- Spring '26 objectives on building blocks and hybrid reasoning.
- Trap
- Thinking it replaces the reasoning engine.
-
Hybrid reasoning
Mixing scripted deterministic steps with model reasoning.
- Remember
- Use scripts where steps must not be skipped.
- Tested as
- Strict sequence scenarios.
- Trap
- Scripting every conversation path.
-
Variables
Session values from context or action outputs.
- Remember
- Feed filters and template expressions.
- Tested as
- Deterministic control questions.
- Trap
- Asking the user for data already known.
-
Filters
Conditions that make topics or actions available only when met.
- Remember
- Stronger than instructions for gating.
- Tested as
- Verification-before-action scenarios.
- Trap
- Using instructions instead.
-
Agent user
Dedicated user a Service agent runs as.
- Remember
- Needs least-privilege permission sets.
- Tested as
- Action failure or overexposure questions.
- Trap
- Granting admin-level access.
-
Employee agent
Agent for internal users in Salesforce and Slack.
- Remember
- Runs with the employee's own access.
- Tested as
- Internal productivity scenarios.
- Trap
- Using it for anonymous customers.
-
Service agent
Customer-facing agent on messaging, email and voice.
- Remember
- Pair with verification and human escalation.
- Tested as
- Self-service scenarios.
- Trap
- Choosing it for internal staff.
-
Messaging for In-App and Web
Channel that embeds agent chat in sites and apps.
- Remember
- Needs an Omni-Channel flow routing to the agent.
- Tested as
- Website deployment questions.
- Trap
- Expecting activation alone to publish the agent.
-
Agent API
REST API for external apps to converse with agents.
- Remember
- Uses client credentials through an external client or connected app.
- Tested as
- Custom app integration scenarios.
- Trap
- Using it where a standard channel works.
-
System context flows
Flows that run without the running user's permissions.
- Remember
- Can bypass agent user restrictions.
- Tested as
- Data exposure questions.
- Trap
- Adding them as actions without review.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Ways to control agent behavior
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Instructions | Natural language guidance | Flexible but probabilistic | Using it for must-always rules |
| Variables and filters | Conditional availability of topics and actions | Deterministic gating | Forgetting to set the variable in an action output |
| Scripted steps | Agent Script deterministic logic | Fixed sequence and branching | Scripting open-ended conversation |
| Topic descriptions | Classification text for routing | Decides which job is chosen | Writing overlapping descriptions |
Employee agent vs Service agent
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Audience | Internal users vs external customers | Determines agent type | Picking by data object instead of audience |
| Running user | Signed-in employee vs dedicated agent user | Determines what actions can access | Assuming Service agents run as the customer |
| Channels | Lightning and Slack vs messaging, email and voice | Determines deployment path | Expecting Slack deployment for customers |
Custom action types
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Flow | Declarative logic and record work | No code, admin-friendly | Running in system context unintentionally |
| Apex | Invocable method with custom logic | Complex processing or callouts | Overusing for simple queries |
| Prompt template | Generated text as an action result | Grounded language output | Using for deterministic updates |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Stronger instructions
- What candidates pick
- Add more forceful wording to instructions to enforce a step.
- Why it looks right
- It is quick and seems clear.
- Why it's wrong
- Instructions are interpreted by the model and can be skipped; filters on variables or scripted steps enforce the rule every time.
Clue to watch for The step must never be skipped.
-
Admin agent user
- What candidates pick
- Give the agent user the System Administrator profile so actions stop failing.
- Why it looks right
- It removes permission errors at once.
- Why it's wrong
- It exposes far more data than the agent needs; grant only the object and field access each action requires.
Clue to watch for An action fails due to access in a Service agent.
-
Topic per question
- What candidates pick
- Create a topic for every kind of question customers ask.
- Why it looks right
- More topics seem more precise.
- Why it's wrong
- Many narrow topics overlap and confuse classification; topics should represent distinct jobs with clear scope.
Clue to watch for Misrouting increases as topics are added.
-
Custom when standard exists
- What candidates pick
- Build a custom Apex action for a common CRM task.
- Why it looks right
- Custom code feels controllable.
- Why it's wrong
- Standard actions and flows already cover common tasks and are easier to maintain and secure.
Clue to watch for The task is a standard record lookup or update.
-
Activation publishes
- What candidates pick
- Activate the agent and expect it on the website.
- Why it looks right
- Activation sounds like go-live.
- Why it's wrong
- Website deployment needs a messaging channel, an Omni-Channel flow routing to the agent and an embedded deployment.
Clue to watch for Customers cannot see the agent on the site.
-
API for everything
- What candidates pick
- Use the Agent API to put the agent on Slack.
- Why it looks right
- APIs seem universal.
- Why it's wrong
- Employee agents deploy to Slack through standard configuration; the Agent API is for custom applications without a standard channel.
Clue to watch for A supported channel is named.
Answer framework
Use these steps when a scenario asks you to design or fix agent behavior.
- Identify the audience and channel to confirm the agent type, Employee or Service, and the deployment path it needs.
- Check that the request routes to the correct topic, adjusting classification descriptions and scope before touching actions.
- For any step that must always happen, use variables with filters or a scripted Agent Script step instead of instructions alone, because the reasoning engine treats instructions as guidance rather than guarantees.
- Choose standard actions first, then custom flows, Apex or prompt templates, with precise names, descriptions and input and output definitions.
- Confirm the running user, agent user or employee, has exactly the access each action needs, and review system-context flows that could bypass sharing rules.
Practice question
Want more like this? The AI-201 practice test has questions for every domain.
Checklist
0 of 14
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Give this section about a third of your study time. Start in a Developer Edition or Agentforce-enabled trial org and build a Service agent with two custom topics, a flow action, a prompt template action and a knowledge action. Then build an Employee agent and compare how each runs, especially which user the actions execute as.
Next, open the new Agentforce Builder and switch between Canvas view and Script view on the same agent. Add a variable set by an action output, a filter that gates another action, and a scripted step that must run first. Test the agent with conversations that try to skip the step. Deploy one agent to an Experience Cloud site through Messaging for In-App and Web so you know every configuration piece. Read the Agent API developer guide overview to understand authentication and session flow.
High-priority ideas are topic design, deterministic control and the agent user. Agent Script and hybrid reasoning are new in Spring '26, so use Salesforce Help and release notes rather than older prep material that predates them.
FAQ
Why is AI Agents the most important section of the AI-201 exam?
AI Agents is 35% of the exam per third-party copies of the current guide, roughly 21 of the 60 scored questions as a planning estimate, and many of its scenarios also draw on prompts and retrieval.
What is Agent Script?
Agent Script is the language behind the Agentforce Builder that defines an agent's topics, actions, variables and deterministic logic. You can edit it visually in Canvas view or as code in Script view.
Is the Atlas Reasoning Engine still tested?
The current objectives describe how an agent works and hybrid reasoning rather than naming Atlas. You should still understand the reasoning engine's classify, plan and act loop.
Are Sales agents still on the exam?
The Spring '26 objective asks only about Employee and Service agents. Earlier outlines included Sales Agent and SDR scenarios, which are no longer listed.
What permissions does a Service agent use?
It runs as a dedicated agent user. That user's profile and permission sets decide what each action can read and change, so grant least privilege.
Do I need to code against the Agent API?
No. You need to know when it fits, such as custom apps or backend systems, and that it uses client credentials authentication through an external client app or connected app.
Which channels should I know?
Messaging for In-App and Web on websites and Experience Cloud sites, email, voice through Agentforce Voice, and Slack and Lightning Experience for employee agents.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 3 knowledge
You have covered agent building blocks, Agent Script, hybrid reasoning, deterministic control, topics and actions, channels, security context, agent types and the Agent API. Practice the largest section under timed conditions.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the AI-201 AI tutor Get any concept explained your way