Agent Builder – Blip | Blip Help

Index

Introduction

The Agent Builder is designed to enable the creation and customization of Artificial Intelligence agents in a visual and intuitive way. With it, you can personalize your agent's behavior, define its knowledge base, set exit conditions for the flow, and integrate external tools.

This tool centralizes all the resources needed for you to build and manage intelligent agents in an intuitive manner.

Creating and configuring agents in the builder

Note: At the end of each example, you will find the configuration JSON file. You can copy it and import it directly into the builder to make any adjustments you want.

How to create an AI Agent block

On the builder screen, click Add block and choose the AI Agent option.

A new AI Agent block is added to the builder.

How to configure the AI Agent block

By clicking on the created AI Agent, you can configure it through the side menu.

Renaming the AI Agent

First, you can rename the block to make it easier to identify. In the following example, the Smart Contact will be a FAQ about AI Agents.

Configuring the Instructions tab

In this section, you can customize your AI Agent to meet your needs. Choose the language model (LLM), set the temperature and maximum number of tokens, enter the agent’s initial instructions, among other options.

Accurate configuration is essential to optimize your agent’s performance.

To configure a model, click the Open model settings button next to the model indication in the Instructions tab. The Configure Instructions screen will be displayed.

In the Model tab, you can choose the version of the model to be used from the dropdown menu.

Still in this tab, you can:

Set the model’s temperature;
Set the maximum number of tokens;
Enable/disable context history (short-term memory).

In the Response tab, it is possible to:

Define whether the agent’s response will be sent to the user in the service channel or stored in a variable.

Set the agent’s response format, with the option to choose from various formats (including a custom format).

After setting the agent’s configurations, simply click Save. If no changes were made, just click Cancel.

Still in the Instructions tab, it is possible to configure instructions for the AI Agent. For each instruction, you can select its level, choosing from: User, Agent, System, Tool, Variable. Let’s go through each one:

After setting the agent’s configurations, simply click Save. If no changes were made, just click Cancel.
Still in the Instructions tab, it is possible to configure instructions for the AI Agent. For each instruction, you can select its level, choosing from: User, Agent, System, Tool, Variable. Let’s go through each one:

System Level

The system level defines the persona and general behavior of the agent. It is the most fundamental instruction, guiding how the AI should behave.
Example: "You are a Python programming expert. Respond clearly and concisely, providing code examples whenever possible."

User Level

The user level is a direct question or command given to the agent during its initial configuration. This allows passing the user’s name to the agent, for example.
Example: "Hello, my name is {{contact.name}}."

Agent Level

The agent level is where instructions on how the agent should respond are provided. This level acts as an intermediary layer between the user and system levels to optimize the agent’s workflow.
Example: You can instruct the agent to create a step-by-step sequence to solve the user’s problem, such as: "Step 1: Understand the user’s problem. Step 2: Generate the Python code. Step 3: Explain the generated code."

Tool Level

The tool level refers to the instructions the AI agent receives to interact with external resources. These tools can be an API, a database, or a search engine.
Example: The agent can be instructed to use a specific API to obtain the temperature, such as "openweathermap."

Variable Level

The variable level refers to specific data that the AI should use, such as names, dates, locations, or any information that changes depending on the context.
Example: If you want to know the temperature of a city, you can use a variable like city, and the AI would use the Tool to get the temperature for the city corresponding to that variable.

In practice:
Add an instruction

Select the instruction level:

Enter the instruction:

In the System and Agent instructions, you will see a shield icon. It provides suggested guardrail instructions that you can add to your agent. Simply copy them and insert them into the instruction. For this example, we will create a new system instruction and add a guardrail for the knowledge base.

For each instruction, it is possible to perform actions to remove or duplicate the instruction using the buttons:

It is possible to move the instruction by clicking and dragging it using the side button (six dots) that appears when you hover over it. This allows you to organize the order of instructions, which is crucial for the agent’s behavior.

If the context history option is enabled in the model settings, it will also appear in the Instructions screen and can be arranged along with the other instructions.

Configuring the Knowledge tab

Moving on to the Knowledge tab, the following is displayed when opening it for the first time:

For the agent to access information, it needs catalogs. These catalogs centralize knowledge from files and URLs, organizing everything around the same topic. To learn more about catalogs, refer to the tutorial on Knowledge Base.

Since your agent has just been created, it does not have any linked catalogs yet. To add one, simply click the Catalogs button or + Add catalog.

Clicking either button opens the Manage Catalogs window.

To create a new catalog, click Create catalog. You will be redirected to the Knowledge Base management screen, where you can create or remove catalogs.

Following our AI Agents FAQ example, we will create a catalog using an XLSX file with questions and answers and a URL containing information about the types of files accepted in an AI Agent’s knowledge base.

Once created, the catalog will be listed in the Manage Catalogs screen in the builder and will be available to link to your agent. For more details on creating catalogs, see the Knowledge Base tutorial.

To link a catalog to the agent, simply check the checkbox and click Save.

Now, in the Knowledge tab, you can see the linked catalog.

By clicking the edit button of the linked catalog, you can manage your files and URLs. This way, you can choose which information the agent should use in its knowledge base.

You will notice that both the XLSX file and the URL are listed. Simply select the ones you want the agent to use and click Save.

Running the Smart Contact

Now is a good time to test the Smart Contact (SC) with the AI Agent. To do this, connect the Start block to the AI Agent FAQ block and publish the SC flow. In the flow test chat, you can interact with the agent.

Configuration JSON (Knowledge bases will need to be added)

Configuring the Exit Conditions tab

To configure the Exit Conditions tab, let's create a new flow. The idea is to build a Blip Smart Contact with three agents:

Plans Agent: Specialized in the plans offered by Blip.

Products Agent: Specialized in available products.

Orchestrator Agent: Responsible for understanding the user’s intent and directing them to the correct specialist agent.

In this scenario, we will also add deterministic blocks to show that it is possible to combine them with AI Agent blocks.

To start, let’s add deterministic blocks that will welcome the user and collect their name. For this, we will copy ready-made blocks from the Block Library and adapt them for our example.

The initial flow looks like this:

Now, you can create the Orchestrator Agent and define its instructions.
In this example, it will have three system instructions:

The first defines its role as the orchestrator.
The second instructs it to address the user by name, using a variable for personalization.
The third is a guardrail to ensure that the orchestrator does not send unnecessary messages to the user.

The knowledge bases for the specialist agents will be the URLs for Blip’s products and plans.
The instructions for each agent will be as follows:

Role Definition: Describes the agent’s area of expertise.

Competitor Guardrails: Prevents the agent from mentioning competitors or similar services.

Tone of Voice Guardrails: Ensures that responses are in Portuguese and maintain a friendly tone.

Grounding Guardrails: Ensures that all information provided is based exclusively on the knowledge base.

User Name: Ensures that the agent addresses the user by name.

Now, just create the agents, configuring them with these instructions and the knowledge base, as seen in the previous steps.

Blip Plans Specialist Agent:

Blip Products Specialist Agent:

Now you can configure the exit conditions. They allow you to instruct the agent on which block the user should be directed to, ensuring that the conversation flow continues correctly.

Let’s start with the Orchestrator. In the Exit Conditions tab, click the + Add routing button.

In Definitions, fill in the name, description, and target block, and save the instruction. Create an exit condition for each specialist agent.

It is also possible to define an exit condition for exception cases.

Now, for each specialist agent, let’s create an exit condition to return to the Orchestrator Agent. This allows the user to go back and switch context.

And this is how the flow looks:

Now it’s time to test!

Configuring the Tools tab

The Tools tab allows you to configure instructions for the agent to interact with external resources. To see a practical example, let’s go back to the Orchestrator Agent.

By clicking Add tool, a list of options will be displayed. For each one, in addition to setting a name, you need to provide a description — which guides the agent on when to trigger the action — and configure the specific data required for its execution.

Let’s use the Set Contact tool as an example, which is used to store user data. Since we have already captured the name, we will use it to fill in the contact information. To do this, simply change the name, provide the description, and map the variable {{nome}} to the corresponding field.

You can add conditions for the contact definition to occur. For example, it is possible to include a validation for the user’s name.

To test the functionality, let’s adjust the instruction we used to pass the user’s name. Instead of using the variable {{nome}}, we will change the instruction to fetch the name directly from {{contact.name}} and verify that the agent can still address the user by name: