Agentforce · Topic 1 of 1
Agentforce for Administrators
Agentforce for admins: agent use cases, topics, instructions and actions, security and agent permissions, prompts and testing in conversation preview.
- 4 min
- Agents combine topics, instructions and actions; actions come from flows, Apex and prompt templates.
- Use AI for language-heavy tasks; use deterministic automation for exact rules.
- Agents only reach what their permissions allow; the Einstein Trust Layer protects data with LLMs.
What it is
Agentforce lets organizations build AI agents that understand requests in natural language and take actions in Salesforce, for employees or customers. Administrators decide where agents add value, configure what they can do, keep their instructions and prompts up to date, and test them before activating changes.
How it works
Building blocks of an agent
| Part | What it is |
|---|---|
| Agent | An AI assistant for a purpose, such as service or sales, for employees or customers |
| Topics | The jobs the agent handles, each with a description and scope; called subagents in Agentforce documentation since April 2026, with no change in function |
| Instructions | Plain-language guidance inside a topic that tells the agent how to behave |
| Actions | What the agent can do, built from flows, Apex, prompt templates and standard actions |
| Prompt templates | Reusable prompts, built in Prompt Builder, that ground AI output in Salesforce data |
When AI is appropriate
AI agents suit tasks with variable language and repeatable outcomes, such as answering common questions, summarizing records or updating simple data. Deterministic tasks with strict rules, such as calculating a price, are better handled by flows or other automation, which an agent can call as an action.
Security and permissions
- Agents act within permissions: an employee agent typically works with the permissions of the user interacting with it, while customer-facing service agents run as a dedicated agent user, created with minimal access, so you grant the extra permissions its actions need.
- Users get access to a custom agent through a permission set's Agent Access settings.
- The Einstein Trust Layer protects data used with large language models, for example through zero data retention with external model providers. Its LLM data masking is turned off for agents.
- If an agent can't complete a task, check that the right topic is chosen, the instructions are clear, the action exists, and the user it runs as has access to the objects and fields involved.
Maintaining prompts and instructions in Agent Builder
- Edit topic descriptions and instructions to improve how the agent chooses and behaves.
- Update or install prompt templates and actions, for example from AgentExchange.
- Agent Builder is now called Agentforce Builder; the exam guide uses the older name.
- Use conversation preview in the builder for light testing: type sample requests and review which topic and actions the agent chose and why. Simulate mode doesn't run actions or check permissions; Live Test mode runs actions and can change data, so test in a sandbox.
- Activate the agent only after testing, and review conversations after launch to find where it struggles.
A worked example: a service agent keeps answering refund questions with general policy text instead of checking the order. The administrator adds a refund topic with instructions to look up the order first, adds a flow action that returns the order status, makes sure the agent user can read orders, and tests in conversation preview before activating the change.
When an agent fails to act, think "topic, instructions, action, permission" in that order.
How the exam asks about this
- "Test how an agent responds before activating a change." Conversation preview in Agent Builder.
- "An agent says it can't access order data." Check the permissions of the user the agent runs as.
- "Calculate a price exactly by company rules." Use a flow and expose it to the agent as an action.
What to remember for the exam
- Agents combine topics, instructions and actions; actions come from flows, Apex and prompt templates.
- Use AI for language-heavy tasks; use deterministic automation for exact rules.
- Agents only reach what their permissions allow; the Einstein Trust Layer protects data with LLMs.
- Maintain instructions and prompts in Agent Builder and test in conversation preview.
- Review conversations after launch to improve the agent.
Adding an action does not give the agent data access. The user it runs as still needs object and field permissions.
Common mistakes
Expecting the agent to work out strict business rules itself. Give it a flow action.
Activating changes without testing. Use conversation preview first.
Granting broad admin permissions to the agent user. Give only what its actions need.
Check yourself
What are an agent's main building blocks?
Topics, instructions and actions, plus prompt templates used by actions.
Where can an administrator test agent changes before activating them?
In Agent Builder, using conversation preview.
An agent can't read case data it needs. What should you check?
The permissions of the user the agent runs as, along with the topic, instructions and action involved.
Topic 1 of 1 in Agentforce
Test yourself on Agentforce
The full Salesforce Platform Administrator question bank has 650 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by Salesforce. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.