Agent Builder Demo Application - Writer AI Studio

Agent Builder

Agent Builder is in beta. Some features are still in development and are subject to change.

When you create a new agent with Agent Builder, it comes with a demo agent that analyzes customer reviews and drafts personalized responses. This walkthrough shows you how the demo agent works and how to modify it to understand Agent Builder’s core concepts.

This walkthrough covers the following topics:

If you are unfamiliar with Agent Builder, review the Agent Builder Overview to learn more about the different components of an agent.

Start an Agent Builder project

To create an agent with Agent Builder, log in to AI Studio and follow these steps:

  1. Click the Build an agent button in the top right corner of the page.

  2. Create a new agent In the modal that appears, select New Agent to create a new agent from scratch.

For common use cases, prebuilt agent templates are available. These include both general-purpose and industry-specific options, all customizable to your specific requirements.

Once the agent is created, a new tab will open in your browser with the Agent Builder interface.

  1. Return to the Agent Builder interface You can get to the Agent Builder interface any time by going to the AI Studio homepage and selecting the agent you created. You’ll see the Configure Deployment page when you select the agent. To get to the edit interface, click the Edit button in the top right corner.

Preview the demo agent

Writer initializes the new agent with a demo workflow that you can use to get started. The demo agent takes a customer review and uses a series of tools to analyze the review and draft a response to the customer.

Agent Builder views

The Agent Builder interface contains four views:

You can switch between views by clicking the tabs at the top of the page.

Preview the agent’s behavior

Before you start editing, preview the agent to see how it works:

  1. Click the Preview tab to see the agent in action.

  2. Copy and paste the example customer review Copy the example customer review under 1. Copy Review and paste it into the 2.Paste Review text input.

  3. Click Draft response. You’ll see a status message in the UI that the agent is generating a response. Once the agent has finished, you’ll see the response in the UI.

  4. Inspect the agent's blueprint Click the Blueprints tab to see the agent’s blueprint. You’ll see green boxes highlighting the path that the agent took through the blueprint. You’ll learn more about the blueprint in the next section.

Tour of the demo agent architecture

The demo agent consists of an interface and a blueprint. The interface and blueprint are connected through the agent’s state and UI triggers.

Agent state

The agent’s state is a core component of the agent’s behavior. It contains values that the UI and blueprint can access and update. You can view the agent’s state from any view by clicking the State explorer icon in the top right of the page.

Blueprints view

The Blueprints view is where you define the agent’s logic and behavior. It’s a flowchart-like interface where you connect blocks of logic. If you previewed the agent, you’ll see green boxes highlighting the path that the agent took through the blueprint. For the sample review, the agent takes the following steps:

  1. UI Trigger The blueprint started with the UI Trigger element. In this case, the UI Trigger is configured to run when the user clicks the Draft response button in the UI.

  2. Set state to show the progress message The agent used the Set state block to set the progress_message value for the agent state to show that it is generating a response.

  3. Classification The agent uses the Classification block to analyze the review to determine the primary focus of the review, from the following: Packaging, Pricing, Quality, Delivery, or Empty. The agent determined the review fit into the Quality category. Classification blocks are built-in blocks that use a Palmyra model to classify the input based on a set of user-defined categories.

  4. Text generation The agent used the Text generation block to have a Palmyra model draft a response to the customer specifically focused on quality.

  5. Set state for the response The agent stored the drafted response using the @{result} environment variable so it could use it in the UI.

  6. Set state to clear the progress message The agent cleared the progress_message value in the state to remove the loading message from the UI.

View another path through the blueprint

To view another path through the blueprint, go back to the Preview view and paste the following:

I like the style but it's a lot more expensive than other brands without any difference in quality. Click Draft response, and navigate back to the Blueprint view. You’ll see that the agent took a different path through the blueprint, this time selecting the Pricing category and drafting a response using a pricing-related prompt.

Interface view

The Interface view is where you define the agent’s appearance. You build the interface with components that you can customize and reference in the blueprint. This agent’s UI has the following components:

The Interface uses Column container components to layout the Review section in a two-column format.

Hidden sections

You won’t see the hidden section for the results in the UI when you first load the agent. You can see that the section is in the UI from the Interface Layers view, which shows all the components in the UI. The hidden section has an icon next to it indicating that it’s not visible.

Click that element in the Component tree to see its settings. Under Visibility, it’s set to only show up once there is a value in the review_response state variable.

Connecting the UI and the blueprint with state variables

To build full-featured agents, you must be able to connect the UI to the blueprint. The agent’s state is the main way to pass data between the UI and the blueprint. The state contains a set of values that the UI and blueprint can both reference and set. You can also use triggers to connect the UI to the blueprint. Here are the places where the demo agent connects the UI to the blueprint and passes data between them:

  1. UI Trigger The Draft response button in the UI is connected to the UI Trigger element in the blueprint. Click events trigger the blueprint to run.

  2. Customer review input The Text area Input component in the UI binds the user’s input to the customer_review state variable. The blueprint uses this value in the Classification block.

  3. Progress message The Message component in the UI displays the value of the progress_message state variable to the user. The blueprint sets this value after the user clicks Generate response.

  4. Review response The two sections in the response with Text components in the UI depend on the value of the review_response state variable.

    1. The blueprint sets this value after the Text generation block runs.
    2. Once review_response contains a value, the section that had the message “The response will be shown here” disappears.
    3. The section that displays the response appears. The text area in that section contains the value of review_response.

  5. Clear progress message Once the review response is generated, the final Set state block in the blueprint sets the progress_message state variable to an empty string to clear the loading message from the UI.

Passing data between blocks

The demo agent also shows how to pass data between blocks with the @{result} variable. Once the text generation block has finished, it passes its result to the next block in the blueprint, which is the Set state block. The Set state block sets the review_response state variable with the @{result} from the text generation block. Learn more about all available variables in Using Data from Previous Blocks.

Modify the demo agent

Now that you’ve toured the Agent Builder interface, try editing the agent’s behavior.

Add a new category to the classification block

This example shows how to add a new category to the classification block to look for and respond to reviews that focus on sizing issues. First, add another category to the Classification block in the blueprint so that the agent can also classify reviews that are about sizing.

  1. Open the Classification block Navigate to the Blueprints view and click the Classification block in the blueprint to open the block settings.

  2. Add a new category In the Categories section, add a new category called Sizing with the value:

    The review relates to sizing issues or satisfaction with sizing.

  3. Add a new text generation block Add a new Text generation block to the blueprint to generate a response to sizing-related reviews:

    1. Click one of the existing Text generation blocks in the blueprint.
    2. Type Command + c (on Mac) or Ctrl + c (on Windows) to copy the block.
    3. Click onto the blueprint canvas and type Command + v (on Mac) or Ctrl + v (on Windows) to paste the block.

  4. Edit the new text generation block Click on the new text generation block you pasted to edit it. Update the following settings:

    • Alias: Draft sizing response
    • Prompt: update the prompt to focus on sizing issues:
   Take the role of a customer success rep and draft an email response to the customer review below that mentions sizing: @{customer_review}
   
   Address the customer's sizing experience. Thank them if the sizing was accurate, or address their concerns if there were problems with sizing.
  1. Connect the new text generation block to the rest of the blueprint Connect the new text generation block to the rest of the blueprint:
    • Drag a line from the purple dot next to Sizing on the Classification block to connect it to the new Text generation block.
    • Drag a line from the Success connection point on the Text generation block to the Set state block.

Now, preview the agent to see the new behavior. Paste the following review into the input field and click Draft response:

The sizing felt off, I'm normally a small and even the XS was too big.

You should see the agent take a different path through the blueprint and generate a sizing-related response.

Add a new UI component

This next example shows how to add a new UI component that you can incorporate into the agent’s blueprint. The steps show how to prompt the agent to add a discount of a certain percentage when generating a response to the customer review. You will:

From the UI view:

  1. Add a new Select input component Add a new Select input component to the UI. Drag it into the section that contains the area for pasting the review text and the Draft response button. Update the following settings in the Select input component’s settings:

    • Label: Discount %
    • Options: 0, 5, 10, 15 (the keys and values should be the same)
    • Link variable under Binding: discount

  2. Move the component above the Draft response button Move the component above the Draft response button by clicking the three vertical dots on the component’s settings menu and selecting Move up. You can also drag the component up in the UI.

    The UI should now look like this:

From the blueprint view:

  1. Edit the Draft packaging response text generation block Click the Draft packaging response text generation block to open its settings. Edit the block’s prompt to read the discount state variable and add a discount to the response if there is one. For example, add the following to the prompt:
    Offer a discount of @{discount} percent on their next purchase.

Do not offer a discount if the value of @{discount} is 0.

Now, preview the agent to see the new behavior. Paste the following review into the input field:

The packaging was awful.

Then, choose a discount amount from the select input and click Draft response. You should see the agent take a different path through the blueprint and generate a response that offers a discount if you set it above 0. You can update the prompts for the other text generation blocks to have the agent offer a discount for the other categories.

Next steps

Now that you’ve interacted with and modified an agent in Agent Builder, you can build your own. Check out the Agent Builder Quickstart to learn how to build a new agent from scratch.