How to Get

Started with
Agentforce
A step by step implementation
guide for creating an autonomous
AI service agent.
SALESFORCE RESEARCH HOW TO BUILD AN AGENTFORCE SERVICE AGENT | 2

Welcome to Agentforce! Let’s get started.

With Agentforce, humans and AI agents drive customer success together. Agentforce agents can
analyze data, make decisions, and take action on tasks like answering customer service inquiries,
qualifying sales leads, and optimizing marketing campaigns. With Agentforce, organizations can
easily build, customize, and deploy their own agents for any use case, across any industry.

This guide explains the process for configuring and troubleshooting an Agentforce Service Agent.

To do so, we will:
• Define the key building blocks of an AI service agent
• Walk you through the steps in building a service agent
• Explain how to monitor and iterate on agent responses
• Summarize a handful of best practices and sticking points

For the latest information about setting up an Agentforce Service Agent, see
Salesforce Help: Agentforce Service Agent. For more information about AI
agents, topics, and actions, see Salesforce Help: Agentforce: Agents and Copilot.
SALESFORCE RESEARCH HOW TO BUILD AN AGENTFORCE SERVICE AGENT | 3

The Building Blocks of an Agentforce Service Agent

An AI service agent is made up of two fundamental building blocks: topics and actions.

Topics are the jobs you want the agent to be able to do. You can think of a topic as a job that can be
achieved with a consistent set of tools (actions) and rules (instructions) for how to do it. Topics are
composed of these key attributes:

• Classification describes what your topic does and the types of user requests that should be
classified into this topic. The agent uses this description to determine when to use your topic in
a conversation. You should write this description in a way that mirrors how customers ask the
questions you want the agent to answer.

• Scope defines the boundaries of the job — what the agent can and cannot do. When you write this,
think of it as a “job description” that will constrain what the agent can do within this topic.

• Instructions provide context and guidelines on how to do the job. These can include
clarifying common terms, which actions to use, when and how to use them, or how to formulate
responses.
SALESFORCE RESEARCH HOW TO BUILD AN AGENTFORCE SERVICE AGENT | 4

The Building Blocks of an Agentforce Service Agent

Here are some examples of topic compositions:

Topic Order Management Login and Access Issues

Classification Answer questions related Help customers with questions

Description to orders, returns, past about issues accessing the site,
purchases, shipping status, challenges logging in, questions
or repairs. about login failing, or finding their
username or password.

Scope Your job is only to answer Your job is only to help customers
questions about a who cannot log in, by resetting
customer’s order status, their password or looking up their
return status, or return and username.
repair policy. Never initiate
or generate an order or
return.

Instructions If a customer asks for the Always clarify what type of device
status of an order or return, (iOS or Android) the customer is
always offer all options to using before using the “Answer
look up the status using Questions with Knowledge”
either their email address, action to retrieve troubleshooting
order date, or the order ID. information, and always include
the device type in the search
query for knowledge.

Promo codes are used for

discounts during checkout, while
registration codes are used to
register and gain access to loyalty
discounts and early access to new
products.
SALESFORCE RESEARCH HOW TO BUILD AN AGENTFORCE SERVICE AGENT | 5

The Building Blocks of an Agentforce Service Agent

Actions are the tools that the Service Agent uses to do a job.
Examples of standard actions available with Agentforce include:

Action Description

Get Case Retrieve the information from a specified case

Create a case Create a net new case

Password Reset Replace an existing password with a new password

Update Contact Replace contact information, like name, phone number, email, etc.

Get Reservation Retrieve reservation details

You can extend actions using:

• Flows: Workflows to automate business processes and collect data in Salesforce

• Apex: Coding language for developers to build custom features for the Salesforce platform
• Prompt Templates: Reusable prompts to generate AI-driven responses based on specific
inputs, use cases, and requirements
SALESFORCE RESEARCH HOW TO BUILD AN AGENTFORCE SERVICE AGENT | 6

Building a Service Agent: Step by Step

In Agentforce, the steps to build a service agent follow a standard path. The key is to start simple.
Then you can iteratively build, test, and repeat until the agent behaves the way you expect.

01 Review the “Reasoning” step to determine if an issue with the instructions in the topic led to
the error.

02 Click the Knowledge tab. Select a library for your agent. (Or, to create a library, click New
Library.) If you select or create a library, you can configure the knowledge fields or file
uploads that you want your data library to index. To base agent responses on all Knowledge
articles and fields, select All Knowledge Records and Fields. Save your changes.

03 Start with standard topics from the Service Agent templates. These are for common use
cases like case management, account management, order inquiry, and more. Standard
topics include basic instructions and actions you can use as a baseline, which you can then
extend and customize to suit your particular needs.

04 From a standard topic, you can:

• Click on New Version
• Customize all fields on the topic
• Add new actions

05 From the standard actions in that topic:

• Click through to the underlying flow, edit it, and save a new version
• The agent action will run the newest saved and active version of the flow
• Click New Version for any action, then edit the action name and description

06 Can’t find a standard topic template for your use case? You can create a new topic for the
job you want the agent to do.
• Start simple: Begin with high-level topics like “Order Management” or “Product
Troubleshooting”
• If you need two sets of instructions to handle different jobs, you can easily split any topic
into two or more topics
SALESFORCE RESEARCH HOW TO BUILD AN AGENTFORCE SERVICE AGENT | 7

Building a Service Agent: Step by Step

07 Provide a detailed Classification Description. For example, what types of requests should
this topic be used to address?

08 Add Scope text. For example, what are the boundaries or the job description of this topic?

09 DO NOT immediately draft a large amount of instructions. The best practice is to start
with a minimal set of basic instructions, ideally just one, and iteratively build them up as you
simulate conversations. Remember: Less is more.

10 Add actions to the topic so the agent has the tools to complete the use case, such as
sending a password reset link, doing a customer information lookup, doing an order
lookup, and so on.

11 Once you’ve added some basic actions, you should have a topic that clearly defines the
job to be done and contains the required actions to do that job. Now it’s time to see
how it works.

12 The Planner is the canvas within Agent Builder where you can test and monitor your agent’s
reasoning and its responses. You can simulate conversations with your service agent in the
Planner’s preview panel by:
• Asking a question or sending a message you expect to trigger that topic
• Stepping through the conversation in the same way you expect customers will interact
with the agent

13 If the service agent does not respond how you want it to, iterate on the configuration of
your topic by adjusting its classification description, scope, or instructions. (Details on this
process are in the next chapter.)

14 Clear the conversation preview and re-test until your service agent works as desired.
SALESFORCE RESEARCH HOW TO BUILD AN AGENTFORCE SERVICE AGENT | 8

Fine-Tuning Your Agent

When a service agent is not behaving the way you
want, how do you know which lever to pull to get it
on the right track?

The key is to analyze the agent’s responses to your

prompts in the Planner canvas, understand why
the service agent is responding the way it is, and
adjust accordingly.
 SALESFORCE RESEARCH HOW TO BUILD AN AGENTFORCE SERVICE AGENT | 9

1 Problem 1: Service Agent Picks the Wrong Topic

This issue occurs when the service agent does not choose the right topic for a customer question.
You can diagnose it by checking the “Select Topic” box in the plan canvas.

To fix this topic, adjust the Classification Description on the agent’s topic(s) to better capture
the types of customer statements that should invoke each topic. Adding key terms, examples, or
common questions that should trigger this topic often helps.

Check the “Classify

Topic” box in the
Planner canvas to
confirm that the
agent chooses the
right topic for a
customer question.
 SALESFORCE RESEARCH HOW TO BUILD AN AGENTFORCE SERVICE AGENT | 10

2 Problem 2: Service Agent Picks the Wrong Action

This issue occurs when the service agent does not invoke the right action when you want it to. You
can diagnose this in the Planner canvas by confirming that, even though it picked the correct topic, it
didn’t run the expected action.

There are two potential levers to address this issue.

• First, make sure the description of your actions accurately reflects what the action is and when it
should be used.

• Second, add instructions to the relevant topic to guide the agent on when to use each action,
or provide it with additional context that can help it interpret common customer statements
(context) that will help in choosing the right action.

When adding an instruction that expresses some sort of condition (“if this, do that” or “after doing
this, then do that”), be as specific as possible.

Keep in mind that instructions rely on the large language model (LLM) for interpretation,
and the LLM lacks knowledge about your specific use case. Any logic unique to your business,
like whether or not a refund should be issued, should be baked into an action in a declarative way.
For example, a flow-based action can assess the criteria for issuing the refund rather than relying
on an LLM to make this decision..

Add an
instruction to
invoke the right
action when you
want it to.
 SALESFORCE RESEARCH HOW TO BUILD AN AGENTFORCE SERVICE AGENT | 11

3 Problem 3: Action Output Is Incorrect or Problematic

In this scenario, the service agent has invoked the correct action, but that action is not returning the
expected output for a given input. Solving this problem requires troubleshooting the action itself. If
the action is based on Flow or Apex, you need to evaluate the Flow or Apex for any issues.

If the issue is with the Knowledge Action, consider the following options:

• Check the query that the service agent passed to the knowledge action. Does the query provide
enough context that it is reasonable to expect the action to find the right answer? If not, consider
adding instructions on how to form effective search queries for that topic.

• Is the action returning irrelevant knowledge? If so, revisit the Knowledge setup for this agent to be
sure only the relevant fields and records are available to the agent.

• Is the action not finding the right article, or is it finding the right article but not extracting the
right information from it? In either case, you may need to explore adjusting your articles to ensure
the answers to common questions are clear in your knowledge base. Q&A articles with non-
overlapping areas of focus and clear titles work the best.

Check the
Knowledge
setup to
troubleshoot
the action.
 SALESFORCE RESEARCH HOW TO BUILD AN AGENTFORCE SERVICE AGENT | 12

4 Problem 4: Right Topic, Right Action, Bad Response

In this scenario, the service agent chose the right topic and ran the right action — but the end
response to the customer did not meet your expectations. For example, the answer to a customer
may correctly share information from an order status lookup action, but not in the tone or format you
want to use with customers.

Adding instructions to your service agent on how to format customer replies, and what style and tone
to use, can solve these issues quickly and easily.
 SALESFORCE RESEARCH HOW TO BUILD AN AGENTFORCE SERVICE AGENT | 13

5 Problem 5: Something Went Wrong

In this scenario, the service agent returned an error message that said:
“I’m sorry, I can’t help with that.”

An AI service agent is made up of two fundamental building blocks: topics and actions:

1. Conflicting or confusing instructions lead the agent to be unsure how to respond.

2. A legitimate bug/error.

A good place to start troubleshooting this issue is to review the Reasoning Step in the Planner
canvas. The LLM validates each agent response before sending to the customer to ensure it does not
conflict with any instructions and that it provides only information from the available instructions and
actions. The Reasoning Step can help reveal if an issue with the instructions in the topic led to the
error.

This is often the toughest problem to troubleshoot. In a worst-case scenario, you may need to
remove instructions from the topic and slowly add them back in until you can figure out the root
cause of the confusion..

Review the
“Reasoning”
step to
determine
if an issue
with the
instructions in
the topic led
to the error.
SALESFORCE RESEARCH HOW TO BUILD AN AGENTFORCE SERVICE AGENT | 14

Additional Tips, Tricks, and Best Practices

Here are some final best practices to keep in mind:

• Be careful with absolutes in instructions. In Agentforce, service agents take “always” or

“never” literally.

• A winning strategy for Q&A with the knowledge action is to add instructions to ask clarifying
questions to narrow the search space, then include the answers to those questions in the
search query.

• Do not rely on the visual order of instructions in the user interface. Be explicit in your
instructions about the conditions that will exist when something should happen. For example,
an instruction that reads, “After that, do XYZ” will not work well. Instead, write: “After searching
knowledge, do XYZ.”

• Treat your agent as a complete novice. Do not expect that it knows anything beyond what
it can retrieve via an action or has available to it in your topic configuration. Define terms in
instructions and use those terms consistently. Providing context via instructions can improve
performance dramatically.

• If all else fails and your agent isn’t working the way you’d like, strip out instructions.
Troubleshooting often begins with reducing the number and complexity of instructions to
reach consistent baseline functionality.

Here are some final best practices to keep in mind:

• Trailhead: Build Your First Agent with Agentforce

• See What Agentforce Can Do for You
 The information provided in this report is strictly for the convenience of our customers and is
for general informational purposes only. Publication by Salesforce, Inc. does not constitute an
endorsement. Salesforce.com does not warrant the accuracy or completeness of any information,
text, graphics, links, or other items contained within this guide. Salesforce.com does not
guarantee you will achieve any specific results if you follow any advice in the report. It may be
advisable for you to consult with a professional such as a lawyer, accountant, architect, business
advisor, or professional engineer to get specific advice that applies to your specific situation.

© Copyright 2024, Salesforce, Inc. All rights reserved.