AgentFlow’s guided setup helps you create an AI agent by describing what you want it to do in your own words.

Based on your use case and available business information, AgentFlow can ask follow-up questions and generate a playbook for your agent. You can then review the playbook, configure how the agent works, test its responses, and publish it to your messaging channels.

In this article, you’ll learn how to:

• Create an AI agent using guided setup
• Describe the conversations you want the agent to handle
• Review and edit the generated playbook
• Set up Knowledge sources
• Configure Context & Response settings
• Configure exit conditions
• Test and improve the agent
• Publish the agent to selected channels

Before you begin

To use this setup experience, you need:

• Access to an AI plan
• Access to the SleekFlow AI module
• Feature Preview enabled for your workspace

Create an AI agent

Step 1: Accessing the setup page

You can follow the steps below to access the AI agent setup page:

1. Click on the  icon to access the “SleekFlow AI” page
2. Click the “Create AI agent” button on the top right corner
3. You will be redirected to the” AI agent setup” page
   
    

Step 2: Describe the agent’s use case

On the agent setup page, describe the customer conversations you want the agent to handle.
 

You can either:

• Enter your own use case in the text field
• Select one of the suggested quick starts

For example:

• Answer product questions and recommend suitable items.
• Help customers check their order status and delivery timeline.
• Qualify new leads before handing them to the sales team.
• Answer appointment booking questions and collect the required customer details.

The more specific your description is, the more relevant the generated playbook is likely to be.

Use a suggested quick start

AgentFlow may display suggested use cases based on the industry and website URL registered for your workspace.

Click a suggestion to use it as your starting point. You can also ignore the suggestions and enter your own use case.

Selecting a suggested quick start follows the same setup process as entering a custom use case.

How AgentFlow uses your business information

This can include:

• Your workspace’s registered industry
• Your workspace’s registered website URL
• Products and services described on your website
• Common customer questions or support information published on your website
• Relevant website pages
• Brand information and tone of voice presented on your website

AgentFlow only uses information it can find from the available workspace details and registered website. It does not use your previous customer conversations to generate the playbook.

If some information is not available on your website, AgentFlow continues with the context it can find and may ask follow-up questions.

Step 3: Answer setup questions

After you submit your use case, you will be redirected to the “Agent configuration” page, AgentFlow may ask follow-up questions to collect additional business context.

AgentFlow only asks questions when it needs more information to generate a relevant playbook.

For each question, you can:

• Select one of the suggested answers
• Select Other and enter your own answer
• Skip the question
• Go back and edit answers within the current round
• Review your answers before confirming them

Open-text answers can contain up to 500 characters per field.

Confirm your answers

After you confirm a round of answers, AgentFlow reviews the new information.

One of the following happens:

• If enough context is available, AgentFlow starts generating the playbook.
• If more context is required, AgentFlow asks another round of questions.

If you skip every question in a round, AgentFlow continues with the information already available. Topics you skip are not asked again.

Step 4: Review and edit the generated playbook

After AgentFlow has enough information, it generates a playbook for the agent.

The generation screen shows the setup progress. When generation is complete, the Playbook canvas opens automatically.

The playbook defines how the agent should think, respond, and act during customer conversations.

Review the Playbook canvas

The Playbook canvas includes:

• An editable agent name
• Generated playbook instructions
• A formatting toolbar
• An Insert action button
• A slash command menu
• Inline action chips
• A configuration summary for Knowledge, Context & Response, Integrations, and Exits

Review the generated content carefully and update any instructions that do not match your intended workflow, business rules, or tone of voice.

Edit the agent name

Click the agent name to replace the generated name with one that is easier for your team to recognize.

The agent name is used to identify the agent inside SleekFlow. It is not necessarily shown to customers.

Edit the playbook

Use the playbook editor to add, remove, or reorganize instructions.

You can use the formatting toolbar to make the playbook easier to scan and maintain.

Your playbook should clearly explain:

• What the agent is responsible for
• What information it should collect
• How it should respond
• Which actions it should perform
• When it should stop handling the conversation

Insert actions into the playbook

Action chips represent actions the agent can perform while handling a conversation.

To insert an action:

1. Click Insert action, or type / inside the playbook editor.
2. Select the action you want to add.
3. Configure the action if required.

An action chip may be:

• Ready to use
• Marked as needing configuration

Set up Knowledge base

Knowledge base determines what information the agent can use when answering customers.

From the Playbook canvas, click “Knowledge” to open the agent’s Knowledge base settings.

You can set up Knowledge base in two ways:

• Use Co-Pilot for guided assistance
• Add and manage Knowledge sources manually

Set up Knowledge base with Co-Pilot

Use Co-Pilot when you are unsure what information the agent needs.

Based on the agent’s use case and the types of questions it is expected to handle, Co-Pilot can help you identify useful information to add, such as:

• Product information
• Frequently asked questions
• Policies
• Service details
• Support documents
• Website content

Review all suggested Knowledge sources before using them. Make sure the information is accurate, current, and relevant to the agent’s purpose.

Set up Knowledge base manually

You can manually:

• View Knowledge sources assigned to the agent
• Browse the Global knowledge base to assign or unassign existing sources
• Upload files
• Index website pages
• Create Custom Q&A
• Add Live Web Search sources

Sources created from the add-source options inside the agent’s Knowledge settings are automatically assigned to the current agent.

Assigned sources can be used by the agent when answering customer questions.

Sources in the global library that have not been assigned to the agent do not display agent-level article previews until they are assigned.

Upload a file

Supported file types include:

• .PDF
• .DOCX
• .XLSX
• .TXT

The maximum file size is 50 MB per file.

Duplicate sources are allowed. Review the assigned sources regularly to avoid giving the agent conflicting or repeated information.

To upload a file:

1. In the “Knowledge base” page, click on the icon on the top right corner
2. Select “Upload files”
3. A side panel will appear
   
4. Drag and drop your files or click “Browse files” to select files from your device
5. Once you have finished uploading, click “Save and process” to continue

Index website content

Use indexed website content when you want the agent to answer from a fixed set of web pages that SleekFlow scans and adds to the Knowledge base.

This option works best for content that:

• Changes infrequently
• Should be reviewed before the agent uses it
• Needs to remain available as a Knowledge source
• Comes from specific pages, such as product pages, FAQs, policies, or service information

When you index a web page, SleekFlow scans the page and processes its content into Knowledge articles that the agent can use when answering customers.

To index a web page:

1. In the “Knowledge base” page, click on the icon on the top right corner
2. Select “Index web pages”
3. A side panel will appear
   

6. Type the URL you would like to index
   1. To scan additional pages under the same URL, select Scan all subpages. SleekFlow can discover up to 200 subpages.
   2. Use “Scan all subpages” when the information the agent needs is spread across multiple pages of the same website. If you only need content from one page, leave this option unselected.
7. After the scan is complete, review the discovered pages and select the ones you want to index. Only add pages that contain accurate and relevant information for the agent.

Add a Live web search source

Use Live web search when the agent needs to retrieve information directly from a website at the time it answers a customer.

This option works best for information that:

• Changes frequently
• Needs to stay current
• Is not practical to maintain as a fixed Knowledge article
• May differ depending on when the customer asks

For example, Live web search may be more suitable for frequently updated availability, schedules, announcements, or other time-sensitive website content.

Unlike indexed website content, Live Web Search does not create static Knowledge articles. The agent searches the configured website when it needs information for a response.

To add a Live web search source:

1. In the “Knowledge base” page, click on the icon on the top right corner
2. Select “Index web pages”
3. A side panel will appear
   
4. In the side panel, enter the website URL you want the agent to search. The URL must begin with http:// or https://.

5. Review the source details and make sure the website contains information relevant to the agent’s purpose.
6. Click “Save and process” to add the source.

The Live web search source is added to the Knowledge base and automatically assigned to the current agent when you create it from the agent’s Knowledge settings.

Because the agent retrieves the content live, its response depends on the information available on the website at the time of the search. Changes to the website do not require you to reindex the source.

Tip: Add only websites you trust. The agent may use information retrieved from the configured website when responding to customers.

Choose between indexed website content and Live Web Search

Use indexed website content when you want more control over exactly which pages and information the agent can use.

Use Live Web Search when keeping the information current is more important than storing and reviewing it as a fixed Knowledge source.

You can also use both options for the same agent. For example, you can index stable product and policy pages while using Live Web Search for information that changes regularly.

Configure “Context and response”

Context and response settings control which customer information the agent can access and how it delivers its replies.

From the Playbook canvas, click “Context & Response”.

Configure “Data access”

Under Data access, select the properties the agent can read while handling a conversation.

You can give the agent access to:

• Contact properties
• Custom object properties

Selected properties become available to the agent as additional conversation context.

For example, you could allow the agent to read a contact’s preferred language, membership tier, or associated order information.

No property selection is required. If you do not select any properties, the agent does not receive additional context from Contact or custom object properties.

Only provide access to information the agent needs to complete its assigned tasks.

Configure the Response engine

The Response engine controls how the agent generates and delivers replies.

Under Response Engine, you can configure:

• Semantic message splitting
• Calculate lead score

Configure Semantic message splitting

Turn on “Semantic message splitting” to let the agent divide long replies into multiple, natural-feeling chat bubbles.

This can make longer responses easier to read and feel more conversational.

Turn this setting off if you want the agent to send each reply as a single message.

Configure Calculate lead score

Turn on “Calculate lead score” to let the agent score each conversation from 0 to 100 based on intent signals.

When this setting is enabled, you can create scoring criteria that reflect the signals most relevant to your business. For example, you can score a conversation based on purchase intent, budget, product fit, or decision timeline.

Configure Exit conditions

Exit conditions define when the agent should stop handling a conversation and what should happen next.

From the Playbook canvas, click “Exits”.

Your agents include two default exit conditions:

• Timeout: Ends the agent’s handling of a conversation after the contact has not replied for the configured amount of time.
• Human escalation: Ends the agent’s handling of a conversation when the contact asks to speak with a human or when the conversation needs to be escalated.

You can expand each condition to review or edit its trigger, description, and follow-up actions.

You can also add more exit conditions based on your workflow.

You can configure exits in two ways:

• Use Co-Pilot for guided setup
• Create and edit exit conditions manually

Set up exit conditions with Co-Pilot

Use Co-Pilot when you are unsure how to translate your workflow into exit conditions.

Describe the outcome you want in plain language. For example:

• Send high-intent leads to the sales team.
• Escalate the conversation if the customer sounds frustrated.
• Stop the agent if the customer asks for a human.
• Follow up if the customer does not reply after one day.

Co-Pilot may ask follow-up questions about your workflow or team structure. It then suggests one or more exit conditions.

Review each suggestion before adding it.

Co-Pilot does not automatically apply suggested exit conditions.

Create an exit condition manually

Use an exit condition to define when the agent should stop handling a conversation and what should happen immediately afterward.

To create an exit condition:

1. In the “Edit agent” page, click “Exits”
   
2. Click “Add exit condition”

3. Fill in the details in the exit condition section:
   
   1. Condition name: Enter a short, descriptive name that explains when or why the agent should exit
   2. Trigger: Select the event or situation that should cause the agent to stop handling the conversation.
      1. Available trigger types include:
         1. Detected sentiment
         2. Time elapsed without a contact response
         3. Goal achieved
         4. Human or escalation requested
         5. System error
         6. Keyword match
   3. Describe: Explain the specific situation that should activate the exit condition.
      1. Be clear about what the agent should look for. The information you enter should match the selected trigger.
      2. For example:
         1. For Detected sentiment: “The customer is frustrated, angry, or dissatisfied.”
         2. For Goal achieved: “The customer has confirmed their budget, preferred product, and purchase timeline.”
         3. For Human or escalation requested: “The customer asks to speak to a person, agent, or support representative.”
         4. For Keyword match: “The customer uses the words refund, cancel, or complaint.”
         5. For Time elapsed without a contact response: “The customer has not replied for 24 hours.”
   4. Follow up action: Follow-up actions define what should happen immediately after the agent exits.
      1. Click Add action to configure an action, such as:
         1. Send a message
         2. Send media
         3. Assign the conversation to a teammate or team
         4. Add a label
         5. Add the contact to a list
      2. You can add up to two follow-up actions to each exit condition. For example, when a customer asks to speak to a person, you can:
         1. Send a message confirming that the conversation is being handed over.
         2. Assign the conversation to the support team.
      3. Follow-up actions are optional. If you do not add an action, the agent stops handling the conversation without taking any further action.

Manage exit conditions

The When to exit page displays all exit conditions configured for the agent.

Each condition card shows:

• The condition name
• The selected trigger type
• A menu for additional actions
• An arrow for expanding or collapsing the condition

Click theicon beside the condition name to rename it.

Click the arrow on the right side of a condition to review or edit its trigger, description, and follow-up actions.

To remove a condition:

1. Expand the exit condition.
2. Click “Delete exit condition”.
3. Confirm the deletion if prompted.

A newly created agent includes default Timeout and Human Escalation conditions.

Keep the following behavior in mind:

• The last remaining exit condition cannot be deleted.
• Disabled conditions remain in the list but are not included in the Playbook configuration summary.
• You can create duplicate or overlapping conditions.
• Overlapping conditions may cause more than one exit workflow to apply, so review them carefully to avoid conflicting follow-up actions.

You can also click “AI Assist” to describe the exit behavior you want in plain language and receive suggested exit conditions.

Step 5: Test the agent

Before publishing your agent, test how it responds to different customer questions and scenarios.

You can test the agent in two modes:

• Chat: Have a test conversation with the agent one message at a time.
• Batch test: Run multiple questions and review the results together.

Use Chat mode to explore a conversation in detail. Use Batch test to check the agent’s performance across a wider range of likely customer questions.

Test the agent in Chat mode

Chat mode lets you interact with the agent in a test conversation without sending messages to real customers.

To test the agent in Chat mode:

1. Click “Chat” in the top-right corner.
   
2. Enter a test message.
3. Send the message and review the agent’s response.
4. Continue the conversation to test how the agent handles follow-up questions and maintains context.

Test a range of scenarios, including:

• Common customer questions
• Follow-up questions
• Incomplete or unclear requests
• Questions that require information from Knowledge sources
• Requests that should trigger an action
• Requests that should cause the agent to exit or escalate

You can also review the available execution details to understand how the agent handled the message. Depending on the response, these details may include:

• The detected intent
• The matched playbook section
• The Knowledge sources used
• The actions triggered
• The agent’s execution path

Use this information to identify whether the agent followed the correct instructions and used the expected source or action.

Run a batch test

Batch test lets you test multiple questions at the same time. This is useful for checking the agent against a broader set of customer scenarios before publishing it.

Click “Batch test” in the top-right corner to open the batch test page.

You can add test questions in two ways:

• Generate suggested
• Write your own

Generate suggested questions

Select Generate suggested to let AgentFlow create test questions based on the agent’s playbook and assigned Knowledge data sources.

Use this option when you want a quick starting set of questions that reflects the agent’s current setup.

Suggested questions are added to the batch test list and run automatically.

Write your own questions

Select “Write your own” to add questions that reflect your actual customer conversations or specific scenarios you want to validate.

Use this option to test:

• Common questions from your customers
• Business-specific rules
• Edge cases
• Questions that previously produced an incorrect response
• Exit or escalation scenarios

After you add the questions, they appear in the batch test list and run automatically.

Review a batch test response

Click a question in the batch test list to open the Evaluate answer panel.

The panel shows:

• The test question
• The agent’s response
• The workflow or response mode used
• Available execution details
• Options for rating the response

Review whether the response:

• Answers the test question correctly
• Uses accurate and relevant information
• Follows the agent’s playbook
• Uses the expected Knowledge sources
• Triggers the correct workflow, action, or exit condition
• Matches the tone and behavior you expect from the agent

You can click “Re-run” at the top right corner to test the same question again using the agent’s latest configuration.

Review whether the response is accurate, relevant, and consistent with the agent’s playbook and Knowledge sources.

Rate the response

Under “Rate the response”, select one of the following ratings:

• Good [G]: The response is accurate, complete, and meets the expected behavior.
• OK [A]: The response is generally acceptable but could be clearer, more complete, or better aligned with the expected behavior.
• Poor [P]: The response is incorrect, irrelevant, incomplete, or does not follow the intended workflow.

You can also enter an Internal note to explain your rating or record what should be improved.

Internal notes are optional and are not shown to customers.

Click “Save” to save the rating and notes.

Improve Poor-rated responses with Co-Pilot

When at least one response is rated “Poor”, click “Review with Co-Pilot” to start a debrief of the Poor-rated test cases.

If some test cases have not been rated yet, a confirmation dialog appears. It shows:

• How many test cases still need a rating
• How many Poor-rated cases will be included in the debrief

You can choose:

• Keep rating: Return to the batch test results and continue rating the remaining responses.
• Review cases: Continue to Co-Pilot using only the cases currently rated Poor.

Only Poor-rated cases are included in the Co-Pilot debrief. Unrated, Good, and OK responses are not included.

Co-Pilot uses the Poor-rated responses and any saved internal notes as context to identify possible issues.

Co-Pilot may:

• Identify patterns across Poor-rated responses
• Ask follow-up questions about the intended behavior
• Suggest changes to the playbook
• Recommend adding or updating Knowledge sources
• Direct you to the relevant agent settings
• Suggest changes to actions or exit conditions

You can review each suggestion and choose whether to accept, modify, or skip it.

After making changes, re-run the affected questions or select Re-test all to check whether the responses have improved.

Filter batch test results

Use the “Status” and “Rating” filters above the question list to narrow the results.

For example, you can filter the list to review:

• Questions that are still running or pending
• Questions that have completed
• Responses rated Poor
• Responses that have not yet been rated

This helps you focus on tests that still require review or improvement.

Manage batch tests

Click Manage to access additional batch test actions:

• Run pending: Run questions that have not completed.
• Re-test all: Run every question again using the agent’s latest configuration. Use this after making broad changes to the agent.
• Reset results: Clear all responses, statuses, ratings, and notes from the current test results.
• Regenerate suggested: Replace the suggested test questions with a new set based on the agent’s current playbook and Knowledge sources. Use this when the agent’s purpose, playbook, or Knowledge sources have changed.

To test a single question again, click Re-run for that result.

Improve the agent based on test results

Review responses rated OK or Poor and identify what caused the issue.

You may need to:

• Clarify or update the playbook
• Add or revise a Knowledge source
• Adjust an action
• Update an exit condition
• Change Context & Response settings

After making changes, re-run the affected questions or select Re-test all to validate the updated configuration.

When at least one response is rated Poor, you can also use AI Assist to review the test results and receive suggestions for improving the agent.

Step 6: Publish the agent

When you are ready to make the agent available to customers, click “Publish” in the agent setup header.

The Publish dialog lets you configure:

• Channels
• Target audience
• Reply hours

Select channels

Under Channels, select the channel instances where the agent should run.

Supported channel groups shown in the Publish dialog include:

Under Channels, select the messaging channel instances where the agent should run.

AgentFlow supports all messaging channels connected (except voice and email at this stage) to your SleekFlow workspace. The available options in the Publish dialog depend on the channels and channel instances currently connected to your workspace.

You must select at least one channel before you can publish the agent.

The “Publish” button remains disabled until a channel is selected.

Select the target audience

Under “Target audience”, choose who the agent should respond to.

You can select:

• All conversations
• Custom audience

When you select “Custom audience”, you can create rules using:

• Label
• Phone
• Keyword

Empty custom audience rules are ignored when you publish the agent.

Review your audience settings carefully to make sure the agent is only deployed to the intended conversations.

Configure Reply hours

Reply Hours control when the agent is active.

If Reply hours are turned off, the agent is always on.

If Reply hours are turned on:

1. Select the days when the agent should reply.
   
   1. Set a “From” time.
   2. Set a “To” time.

Overnight schedules are supported. For example, you can configure the agent to reply from 22:00 to 06:00.

Complete publishing

After you finish configuring the agent:

1. Review the selected channels.
2. Review the target audience.
3. Review the reply hours.
4. Click “Publish”.

Publishing applies the selected channel, audience, and reply-hour settings.

Clicking “Cancel” closes the dialog and discards any unpublished changes made inside it.