Agentforce Specialist AI agents explained
AI Agents is worth 35% of the Salesforce Agentforce Specialist exam — around 21 of the 60 scored questions and comfortably the largest section. It covers how agents are built, how they behave, where they run and what they are allowed to do.
How an agent works
The foundation: an agent receives input, reasons about what is being asked, selects a topic, and executes actions to fulfil it.
The outline names the basic building blocks of agent script, and hybrid reasoning with its components and benefits. Hybrid reasoning combines the flexibility of a model with deterministic control, which is the central design tension of the whole product: you want an agent that can handle unanticipated phrasing but always calculates a discount the same way.
Agent Script is authored in two surfaces — Canvas and Script View — and the outline asks how it functions in each. Know that they are two views onto the same thing, and what each is suited to.
Topics and actions
The configuration layer, and the most testable part of the section:
| Building block | What it is |
|---|---|
| Standard topics | Provided by Salesforce for common scenarios |
| Custom topics | Defined by you for your business cases |
| Standard agent actions | Prebuilt operations the agent can perform |
| Custom agent actions | Operations you define |
A topic is a job the agent can do; actions are the steps it takes to do it. Questions present a use case and ask which combination fits, and the recurring judgement is standard versus custom: use standard where it exists. Building a custom action that duplicates a standard one is the wrong answer.
Deterministic behaviour
Explicitly called out: managing deterministic behaviour using filters, variables and template expressions.
This is the answer whenever a requirement says always, never, exactly or the same every time. A model asked nicely to be consistent is not consistent. Filters constrain what is considered, variables hold known values, and template expressions produce fixed output.
Any scenario about a calculation, a legally required phrase or a fixed format is a deterministic-mechanism question — never a prompt-wording question.
Channels
Connecting agents to digital experience, email, voice and Slack. Know that an agent is built once and surfaced across channels, and that each channel has implications for how users interact and what context is available.
The security context
Explain the security context in which the agent is actually running, and how it impacts agent action execution.
This is the objective that most often surprises people in practice. An agent executes actions as some identity, and that identity’s permissions determine what it can actually do — regardless of what the topic says it should do. An agent configured to update a record cannot do so if the running context lacks permission.
Expect scenarios where an agent behaves correctly in testing and fails in production because the security context differs.
Employee versus Service agents
Given a scenario, identify when to use an Employee or Service agent. The distinction is audience: internal staff versus external customers. That drives the channels, the data, the guardrails and the tone.
The Agent API
Given a scenario, identify when it’s appropriate to use Agent API. The answer is when an agent must be reached programmatically or embedded in an experience outside Salesforce’s standard channels.
Sample questions
Question 1. An agent must quote a shipping fee that is legally required to be stated in identical wording every time. How should this be configured?
- A. Add an instruction to the topic asking the agent to use consistent wording
- B. Select a different model for the agent
- C. Use a template expression with variables to produce the fixed wording
- D. Remove the shipping fee capability from the agent
Show answer
Answer: C
A requirement for identical output every time calls for deterministic mechanisms such as template expressions and variables rather than relying on the model’s phrasing. Instructing the agent in a topic does not enforce wording, a different model does not guarantee consistency, and removing the capability fails the requirement.
Question 2. An agent updates case records correctly in a sandbox but silently fails to do so in production. Topic and action configuration are identical. What should be investigated?
- A. The security context the agent runs in, and its permissions in production
- B. The agent’s channel configuration
- C. The Data Library attached to the agent
- D. The chunking strategy used for grounding
Show answer
Answer: A
Action execution depends on the security context the agent runs in, and differing permissions between environments explain identical configuration behaving differently. Channel configuration affects where the agent appears, the Data Library affects grounding, and chunking affects retrieval quality.
Question 3. A company needs an agent that answers customer questions on its public help site, and a separate agent that helps staff find internal HR policies. What should be built?
- A. One agent serving both audiences, with a filter on the topic
- B. Two Employee agents, one per audience
- C. An Employee agent for customers and a Service agent for staff
- D. A Service agent for customers and an Employee agent for staff
Show answer
Answer: D
The audiences differ, so a Service agent serves external customers on the public site while an Employee agent serves internal staff, each with appropriate channels, data and guardrails. One agent for both risks exposing internal content externally, and the other pairings invert the roles.
What to practise
Build one agent with a standard topic and one custom topic, give it a custom action, and add a filter so one behaviour is deterministic. Then surface it on a second channel.
Finally — and this is the exercise that pays off — run it as a user with fewer permissions than yourself. Watching an action fail because of the security context teaches that objective in a way no reading does.