Build an Agent Using Agentforce Vibes

Vibe code an agent by having a natural language conversation that describes what you want your agent to do. Then watch Agentforce Vibes generate the code.

Agentforce Vibes is enabled by default in supported editions. For developer environment prerequisites, see Set Up Agentforce Vibes.

To build an Agentforce agent, you must first enable Einstein and Agentforce in your org.

Don’t have a development org? Sign up for a Developer Edition org using this form.

Log in to your development org.
From Setup, in the Quick Find box, enter Einstein Setup, and then select Einstein Setup.
Make sure that Einstein is enabled. If it isn’t, select Turn on Einstein.
In the Setup Quick Find box, enter Agent, and then select Agentforce Agents.
Make sure that Agentforce is enabled. If it isn’t, click On.

If you just enabled Agentforce for the first time, refresh the page to see the New Agent button.

Agents are goal-oriented, autonomous AI assistants that perform tasks and business interactions. They can initiate and complete a sequence of tasks, handle natural language conversations, and securely provide relevant answers drawn from business data.

Agentforce agents use Agent Script as their blueprint. Agent Script is a language that combines the flexibility of natural language vibe coding with the reliability of programmatic expressions for handling business rules. Agent Script gives you all the advantages of access to a large language model (LLM), while also providing ways to add more deterministic behaviors to your agent.

Agentforce Vibes provides a welcome page to help you get started building an agent.

From VS Code, open an existing DX project and authorize an org using the SFDX: Authorize an Org command from the Command Palette.

Agentforce Vibes requires that you authorize an org before you can use any of its features.
From the Command Palette, run Agentforce: Show Welcome Page.
Click Agents.

You now have two options to build an agent:
- Create a Project: Creates a DX project that contains a sample agent called Local Info Agent. Enter the project name and location, then click Create Project. If prompted, click Yes, I trust the authors. See Create a DX Project with a Sample Agent for what to do next.
- Start with a prompt: Start vibe coding right away using a natural language prompt. Read the short descriptions of the sample prompts, click one to copy it to the chat interface, then send it to start generating code. See Start From a Prompt for what to do next.

One way to learn about building Agentforce agents is to create a Salesforce DX project that contains a sample agent and explore its features. You can create a DX project using either of these options:

From the Agentforce Vibes welcome page.
Run the SFDX: Create Project command from the Command Palette and select the Agent project template.

When the DX project is created, the VS Code Explorer opens showing the files in the DX project. The project contains a sample agent called Local Info Agent that you can, for example, embed in a resort’s website. The agent provides local weather updates, shares information about local events, and helps guests with facility hours.

The agent demonstrates:

Three types of subagents (Invocable Apex, Prompt Template, and Flow).
Mutable variables.
Flow control with available when.
Deterministic branching with if/else in reasoning instructions.

You must authorize your development org in your new DX project so you can explore the sample Local Info Agent.

Run the SFDX: Authorize an Org VS Code command from the Command Palette.
Click Production, unless you’re using a sandbox or an org on a custom instance URL.
Enter the alias for your org.
In the browser that opens, log in to your org.

When you see Authentication Successful, you can close that browser window.

Agentforce agents use an Agent Script file as their blueprint. To vibe code an agent, you vibe code its Agent Script file. Agent Script files are part of the AiAuthoringBundle metadata type.

In the VS Code editor, open the force-app/main/default/aiAuthoringBundle/Local_Info_Agent/Local_Info_Agent.agent file.
Open the Agentforce Vibes panel from the Activity Bar.
In the chat interface, enter prompts to learn how the Local Info Agent is coded. Here are a few sample prompts to get you started:
- What does the Local Info Agent do?
- What Apex classes does this agent use?
- Does the agent use flows?
Enter prompts to change how the sample agent works.

You can start with a simple prompt, such as:

Add a subagent to the Local Info Agent that helps guests plan their day at the resort.

But Agentforce Vibes works best with detailed prompts. For example:

Add a subagent to the Local Info Agent that helps guests plan their day at the resort. The subagent should check the current weather and look up facility hours, then suggest a schedule of activities based on the results. Make sure to create an agent user for this task. Use existing permission sets if possible. Validate the agent’s behavior before completing this task. If you can’t find the Local Info Agent, let me know and suggest that a new project be created using the ‘agent’ project template.

See Start From a Prompt for more examples.

Previewing an agent refers to having a conversation with it, which acts as a first-line test. You can preview an agent programmatically with an Agentforce Vibes prompt, or interactively in the Agentforce DX panel.

If you haven’t yet deployed the Apex classes, flow, or prompt templates to your org, your preview is simulated and the Local Info Agent mocks the answers to your questions. After you deploy, you can preview in live mode, in which the agent uses the actual Apex classes, flows, and prompt templates.

To vibe code a preview, run prompts like this in the chat interface:

Preview the Local Info Agent.
Run validation tests on the Local Info Agent. Make sure you test all the subagents.

When the preview completes, Agentforce Vibes displays a summary of what it did and where you can find the trace files to inspect details about the conversation.

To interactively converse with the Local Info Agent:

In the VS Code editor, open the force-app/main/default/aiAuthoringBundle/Local_Info_Agent/Local_Info_Agent.agent file.
Right-click the Local_Info_Agent.agent file and select AFDX: Preview This Agent.
In the Agentforce DX panel that opens, click Start Simulation.
Enter a question in the chat interface at the bottom, such as:
- What can you help me with?
- What’s the weather like?
- What are some local events?
The chat interface in the Agentforce DX panel for previewing agent behavior is different from the Agentforce Vibes chat interface.

This table shows the key Local Info Agent metadata files in your DX project. All files are in the force-app/main/default directory under their associated metadata directory.

Component	Type	Purpose
`Local_Info_Agent.agent`	Agent Script	The agent definition, including subagents, reasoning, variables, and flow control.
`CheckWeather`	Apex Class	Invocable Apex. Checks current weather conditions for a given location.
`CurrentDate`	Apex Class	Invocable Apex. Returns the current date for use by the agent.
`WeatherService`	Apex Class	Provides mock weather data for the resort.
`Get_Event_Info`	Prompt Template	Retrieves local events.
`Get_Resort_Hours`	Flow	Returns facility hours and reservation requirements.
`Resort_Agent`	Permission Set	Agent user permissions (Einstein Agent license).
`Resort_Admin`	Permission Set	Admin or developer Apex class access.
`AFDX_Agent_Perms`	Permission Set Group	Bundles the agent user permissions for assignment.
`AFDX_User_Perms`	Permission Set Group	Bundles the admin user permissions for assignment.

Vibe code right away by entering natural language prompts in the Agentforce Vibes chat interface. The welcome page includes sample prompts — select a description and Agentforce Vibes copies the full prompt to the chat interface. You can also copy and run any prompt from this table directly.

For complex agent builds, enable Plan Mode before prompting. Agentforce Vibes analyzes your request, asks clarifying questions, and produces a structured plan you can review and refine before approving execution. See Plan Mode.

Description	Prompt
Creates a design document (Agent Spec) for an existing agent.	Load any skills related to Agent Script, then find existing `.agent` files in my project. Let me know what they are and offer to produce an Agent Spec for any agent I’m interested in learning more about. The Agent Spec should document the agent’s design — subagents, actions, variables, gating logic, and backing logic dependencies.
Extends an agent by adding a new ‘daily planner’ subagent.	Load any skills related to Agent Script, then add a new subagent to the Local_Info_Agent.agent that helps guests plan their day at the resort. The subagent should check the current weather and look up facility hours, then suggest a schedule of activities based on the results. Make sure to create an agent user for this task. Use existing permission sets if possible. Validate the agent’s behavior before completing this task. If you can’t find the Local_Info_Agent.agent, let me know and suggest that a new project be created using the ‘agent’ project template.
Creates an employee-facing agent for searching Salesforce accounts.	Load any skills related to Agent Script, then create a new `.agent` file for an employee agent called Account_Lookup_Agent. It should have a single subagent that lets employees search for accounts by name or list the first 10 accounts in the org sorted by name.

To continue on your agent development trek, check out the Build Agents with Agentforce DX section in the Agentforce Developer Guide, which shows you how to:

Author an agent, which involves generating an authoring bundle, coding the Agent Script file, and publishing the agent to your org.
Preview and debug an agent.
Test an agent.