Agent Component: Complete configuration guide
Learn how to quickly create your first Agent and then explore the full set of configuration options step by step.
Overview
Create your own AI Agent with the Agent Component. This component abstracts the APIs of major LLM providers so you can easily perform tasks such as text classification, information extraction, summarization, content evaluation, and more — all inside your Digibee pipelines.
This guide walks you through:
Understanding each configuration option
Exploring advanced features like tools, retrieval, guardrails, and mock responses
Tip: You can find more real-world examples of the Agent Component in the Use Cases section.
AI-assisted fields
Fields marked with in this documentation support AI-assisted content generation. In the platform, these fields display a sparkle button () in the upper-right corner. Use it to generate or edit field content from a plain-language description, without having to write it manually.
To use the assistant:
Click the sparkle icon (). A panel opens where you can describe what you need.
Enter your description and click Generate.
When generation is complete, you can choose one of the following options:
Apply: Replaces the current field content with the generated result.
Try again: Returns to the prompt to repeat the same description or test a new one.
Cancel: Discards the result and keeps the current field content.
Applying a generated result overwrites whatever is currently in the field.
Basic configuration
With the two steps below, your Agent is ready for simple and direct tasks.
Select the model
Choose the LLM provider and model your Agent will use. Supported providers:
Anthropic Claude
DeepSeek
Google Gemini
OpenAI
Azure OpenAI
Custom Model
Azure OpenAI (Private)
Amazon Bedrock (Private)
Or select Custom Model to personalize the model.
After selecting the provider, click the gear icon (⚙) to open the Model Configuration page. Complete the fields below.
Account Configuration
These settings define how Digibee authenticates with the selected LLM provider. If no account is selected, the connector will not operate. Some providers also require a custom Endpoint, depending on how they expose their models.
Private providers — such as Azure OpenAI (Private) and Amazon Bedrock (Private) — always require a custom endpoint. When you select one of these providers, make sure you provide the correct endpoint for your environment.
For Amazon Bedrock, additional configuration is required:
Select the AWS-V4 account type, the authentication method supported by Bedrock.
Because Digibee does not offer predefined models for Bedrock, you must enable Custom Model and enter the ARN (Amazon Resource Name) of the model you want to use from your AWS account.
Account
The account used to authenticate with the LLM provider. It must be previously registered in Accounts. Supported types: Secret Key or AWS-V4 (for Bedrock).
Account
❌
N/A
Certificate
The account used to enable secure HTTPS communication with external endpoints. It must be previously registered in Accounts. Supported type: Certificate Chain.
Account
❌
N/A
Endpoint
Custom endpoint for the selected provider. Required for private providers (such as Azure OpenAI or Amazon Bedrock).
String
✅
N/A
Allow Insecure Connections
Allows the client to accept insecure SSL connections. Recommended only for testing or internal environments.
Boolean
❌
false
If you select a private model, you must also configure the Custom Model Name. If you choose to fully customize the model, you must set both the Custom Model Name and a compatible provider.
Model Parameters
The parameters available depend on the model you selected. Each field appears only if the model supports it. Parameters not supported by the selected model are hidden automatically.
Temperature
Controls creativity. Lower values produce more focused responses; higher values produce more diverse ones.
Top P
Controls variability using probability thresholds. Lower values produce more focused responses.
Max Output Tokens
Sets the maximum length of the response.
Reasoning Effort
Controls how much the model reasons before responding.
Thinking Budget
Sets the maximum number of tokens the model can use for its internal reasoning process.
Include Reasoning
When enabled, the model's internal reasoning trace is returned alongside the final response.
Verbosity
Controls the level of detail and length of the response.
Advanced Parameters
The remaining supported parameters are grouped under Advanced parameters, collapsed by default. The section header shows a badge with the number of advanced parameters configured — for example, Advanced parameters · 2 configured.
Sampling tuning
Top K: Limits the number of token candidates considered at each step.
Min P: Filters out tokens below a minimum probability relative to the top token.
Top A: Filters tokens based on their probability relative to the top token's probability squared.
Repetition control
Frequency penalty: Reduces repetition of identical words.
Presence penalty: Encourages the model to introduce new topics.
Repetition penalty: Applies a direct penalty to tokens already present in the output.
Output control
Tool choice: Controls whether and how the model uses available tools.
Stop: Defines sequences that signal the model to stop generating.
Seed: Sets a fixed value for deterministic outputs; identical inputs will return the same response.
Debug / observability
Logprobs: Returns log probabilities for output tokens.
Top logprobs: Returns the N most likely tokens and their probabilities at each position.
Logit bias: Adjusts the likelihood of specific tokens appearing in the output.
Custom Parameters
Use these options to fine-tune the model’s performance and resilience. You can modify them if necessary or keep the default configuration.
Custom Parameters
Additional custom parameters to include in the model API request.
Key-value pair
❌
N/A
Timeout
Maximum allowed time (in ms) for the operation to complete.
Integer
✅
30000
Max retries
Number of retry attempts after a failure.
Integer
✅
3
Response Format
Define how the model should structure the output. When supported by the provider, setting a JSON Schema is recommended to ensure more precise and consistent responses.
Use JSON Schema
Enables providing a JSON Schema to enforce response structure.
Boolean
❌
false
JSON Schema definition
The Schema defining the expected structure.
JSON
❌
N/A
Use JSON Mode
Enables sending a JSON example to guide the response.
Boolean
❌
false
JSON definition
The JSON example the model should follow.
JSON
❌
N/A
Use Full Response Mode
Enables additional metadata in the output, including sources and toolExecutions.
Boolean
❌
N/A
Configure the messages
Messages define how the model behaves and how it interprets your prompt.
System message : Describes rules, tone, and general behavior.
Example: “You are a senior data analyst…”
User message : The actual prompt. You can inject pipeline data using Double Braces.
Example: “Summarize the content below…”
After adding your messages, you’re ready to test the Agent.
Every execution consumes tokens from your provider. To reduce token usage when testing the full pipeline, consider adding a mock response.
Advanced configuration
Optional settings let you extend and customize the behavior of the Agent. These are not required, but they are useful when you need integrations, external context retrieval, validation, or controlled testing.
Add tools
Tools expand what the Agent can do. Use them when the Agent needs to:
Call APIs
Access internal systems
Trigger other pipelines
Fetch external web content
If you're only generating text or doing classification, you can skip tools.
Here’s how to configure the supported tool types.
MCP Server
A Model Context Protocol (MCP) server handles communication between the AI model and source systems, enabling the LLM to securely access external tools and data.
Add MCP Tools From Registry
If you have pipelines deployed with the MCP Server Trigger in Test or Prod, they will be available in the MCP From Registry tab.
In this tab, you can view:
The pipeline name
The authentication type
The configured tools and their respective parameters
The version of the deployed pipeline
MCP pipelines currently used as tools in the component (identified with the In use tag)
To add an MCP pipeline as a tool in your Agent Component, select the desired pipeline and click Add Tool. You will be redirected to the Add MCP Tools tab to complete the configuration.
Add MCP Tools
Use this tab to configure the MCP Tool.
If the MPC pipeline was selected through the MCP From Registry tab, some parameters are automatically pre-filled:
Name: Pre-filled with the pipeline name.
Server URL: Pre-filled with the pipeline endpoint.
Headers: If an authentication method is configured, the Key field is automatically populated according to the authentication type:
If API Key is enabled, the header is
apiKey.If Token JWT is enabled, the header is
Authorization.If External JWT is enabled, the header is
externalJWT.
You must manually provide the corresponding Value.
Refer to the parameter descriptions below for more details about each configuration option.
Name
Identifier for this MCP server.
String
✅
N/A
Server URL
Endpoint that receives the request.
String
✅
N/A
Account
The account to authenticate with. Must be configured in Accounts. Supported type: Secret Key.
Account
❌
N/A
Custom Account
A secondary account, referenced via Double Braces, for example {{ account.custom.value }}.
String
❌
N/A
Certificate Account
The account used to enable secure HTTPS communication with external endpoints. It must be previously registered in Accounts. Supported type: Certificate Chain.
Account
❌
N/A
Headers
HTTP headers to include in requests.
Key-value pairs
❌
N/A
Query Parameters
Query parameters to include in the request.
Key-value pairs
❌
N/A
Available Tools
This field is available only when an MCP pipeline is selected in the MCP From Registry tab.
It displays the tools configured in the selected MCP pipeline. Enable the tools you want to make available to the Agent Component.
Boolean
❌
N/A
Content Retrieval
Enable Retrieval-Augmented Generation (RAG) so the model can use external context stored in a vector database.
Embedding Model
An embedding model converts text into vectors that capture its meaning. Choose a provider and model that match your architecture. Supported options:
Local (all-MiniLM-L6-v2)
Max Results
Maximum number of vectors returned from the similarity search.
Integer
❌
N/A
Min Score
Minimum similarity score (0.0 to 1.0) for a result to be considered relevant. Higher is more restrictive.
String
✅
0.7
OpenAI
Embedding Account
The OpenAI account for embeddings. Supported types: Basic, Secret Key.
Select
❌
N/A
Embedding Model Name
Name of the embedding model, for example text-embedding-3-large.
String
✅
N/A
Vector Dimension
Number of dimensions in the embeddings. Must match your vector store’s config.
Integer
❌
N/A
Max Results
Maximum number of similar vectors to return.
Integer
❌
N/A
Min Score
Minimum similarity score for a result to be considered. More restrictive when higher.
String
✅
0.7
Hugging Face
Embedding Account
Hugging Face account credentials. Supported: Basic, Google Key.
Select
❌
N/A
Embedding Model Name
Name of the embedding model (for example, sentence-transformers/all-mpnet-base-v2)
String
✅
N/A
Vector Dimension
Number of dimensions in the embedding vector. Must align with your store.
Integer
❌
N/A
Wait for Model
Whether the system should wait for the model to load (true) or return an error if it’s not ready (false).
Boolean
❌
true
Max Results
Maximum number of vectors returned.
Integer
❌
N/A
Min Score
Minimum similarity score for relevance.
String
✅
0.7
Vector Store
A vector store is a database optimized to store and retrieve embeddings. You can connect to:
PostgreSQL (PGVector)
Host
Hostname or IP address of your PostgreSQL server.
String
✅
localhost
Port
Port number for the Postgres connection.
Number
❌
5432
Database Name
Name of the database containing the vector table.
String
✅
N/A
Vector Store Account
The Postgres account with credentials.
Select
❌
N/A
Table Name
Name of the table where embeddings are stored.
String
✅
embeddings
Neo4j
Database Name
Name of the Neo4j database used for vector storage.
String
✅
N/A
Vector Store Account
Credentials for connecting to Neo4j.
Select
❌
N/A
Index Name
Index that holds the embedding vectors.
String
✅
N/A
URI
URI used to connect to the Neo4j instance.
String
✅
N/A
Node Label
Label for nodes storing embeddings.
String
✅
Document
Embedding Property
Node property that holds the embedding vector.
String
✅
embedding
Text Property
Node property that stores the original text or document.
String
✅
text
Add files
To include external documents (PDFs, images, among others) as context:
Upload them using a storage connector (for example, S3, Azure Blob, Digibee Storage)
Click the plus (+) button.
Enter the filename (for example,
bankslip_1.pdf) or reference it with Double Braces.
The content will be injected into the prompt as additional context for the model.
Setup guardrails
Use guardrails when your inputs may contain sensitive data (like PII), or when you must strictly control what is sent to the LLM.
Click the gear icon (⚙) next to Guardrails to open the Guardrail Settings.
Check the validations you want, or choose Select All to enable all options.
Mask on detection
Masks detected PII before sending to the LLM. If disabled, the execution fails with an error.
CNPJ detection
Detects CNPJ (Cadastro Nacional da Pessoa Jurídica) patterns in the User Message, such as 00.000.000/0000-00. It doesn’t validate the numerical logic.
Credit Card detection
Detects credit card number patterns in the User Message, ranging from 13 to 19 digits. It doesn’t perform checksum (Luhn) validation.
IP detection
Detects IPv4 and IPv6 address patterns in the User Message. It doesn’t validate network or range correctness.
Datetime detection
Detects datetime patterns in the User Message, such as 2025-11-05, 05/11/2025, or Nov 5, 2025. It doesn’t validate date logic.
IBAN code detection
Detects IBAN (International Bank Account Number) patterns in the User Message, such as GB82WEST12345698765432. It doesn’t perform checksum validation.
CPF detection
Detects CPF (Cadastro de Pessoa Física) patterns in the User Message, such as 000.000.000-00. It doesn’t validate the numerical logic.
Email detection
Detects email addresses in the User Message, such as user@example.com. It doesn’t validate domain or address existence.
Crypto Wallet Address detection
Detects crypto wallet address patterns in the User Message, such as Bitcoin addresses starting with 1 and containing 26–35 base58 characters. It doesn’t validate the checksum or blockchain type.
Phone number detection
Detects phone number patterns in the User Message, such as (11) 99999-9999 or +55 11 99999-9999. It doesn’t validate carrier or region.
URL detection
Detects URL patterns in the User Message, such as http://example.com or https://www.example.org/path. It doesn’t verify link validity.
On the same page, you can also:
Enable JSON Schema validation for responses. The output is validated against the schema you provide. If it’s invalid, the Agent sends a reprompt requesting the correct format. If the model still returns an invalid response, the execution ends with an error.
Define custom regex patterns to validate the input. Provide both the Pattern Name and the Regex Pattern.
Test your Agent
Use the Test Panel on the right side of the page to validate your configuration before running the full pipeline.
For detailed guidance on how to perform tests within the Agent Component and explore all available technical details, see the Build Agent tests: Datasets, Test Cases, and Evaluations documentation.
Each successful test will consume tokens, and the number of tokens used is shown in the output. If you want to test the entire pipeline with a specific Agent Component response, you can mock this response during the pipeline run to avoid AI usage costs.
Version the component
The Agent Component supports optional versioning. You can apply the Agent directly to the pipeline or save versions to track and reuse different configurations over time.
You can start versioning the component only after the pipeline is saved.
Without versioning
By default, the component opens in Draft mode. If you choose not to use versioning, configure the component and click Confirm in the upper-right corner to apply the configuration directly to the pipeline.
Only the current configuration is stored, and no version history is created.
With versioning
Versioning allows you to save, organize, and reuse different configurations of the Agent Component.
To create or manage versions:
After making changes to the component, click Save next to the Draft indicator in the upper-left corner and select Create new version.
In the modal, select Create new version or Update existing version, depending on whether versions already exist.
Optionally, add tags to help identify the version, such as the model name (for example, OpenAI – gpt-4o).
Click Save.
Each saved version receives a version number, such as v1, along with any defined tags. Once created, versions cannot be deleted.
To apply a saved version to the pipeline, click Confirm. The version currently applied is marked with the On pipe tag.
Saving and applying changes to a version
After updating the component configuration, you can choose how to proceed:
Save (upper-left corner): Saves the changes to a version and keeps the component open for further editing.
Save and Confirm (upper-right corner): Saves the changes to a version and immediately applies it to the pipeline.
You can access any version you have created at any time from the upper-left corner to load, review, or compare previous versions.
Deploying a version
When deploying a pipeline that includes an Agent Component with versions, the deployment automatically uses the version currently applied to the pipeline.
Tip: To confirm which version is applied, open the component and check which version is marked with the On pipe tag.
After a version is deployed, it can no longer be changed and is marked with the Locked tag. Other versions that were not deployed remain editable and can be updated normally.
Configure general settings
These parameters influence how the component behaves within your pipeline, not how the LLM operates.
They can be accessed on the Settings tab.
Configure step
Step Name: The name displayed in your pipeline.
Alias: An alias you can use to reference this connector’s output later via Double Braces. Learn more.
Configure mock response
Create a mock response to test your pipeline without sending real requests to an LLM.
This is helpful when you want:
Deterministic testing
To avoid usage costs
Mocks substitute live model calls during pipeline execution.
Steps to create a mock:
Click Create mock response.
Provide a Name.
Enter the JSON of the mock response in the JSON Response field.
Activate Set as active to use this mock for tests.
Click Create mock response to save.
Setup error handling
Enable Fail On Error if you want the pipeline to stop on failure. Otherwise, the connector continues and returns { "success": false } inside the output.
Document the Agent Component’s usage
Use the Documentation tab to record:
Use cases
Business rules
Required inputs
Output examples
You can format all of this in Markdown.
FAQ
How can I test and experiment with my prompts?
Use the testing panel located on the right side of the connector’s configuration form. For detailed guidance, see the Testing your Agent topic.
Can I use data from the previous connectors?
Yes. You can use Double Braces expressions to reference data from previous connectors and include it in your prompt.
How is sensitive data handled?
The connector doesn’t redact or filter payload data. We recommend following the same data handling practices used with other connectors.
Can I chain multiple LLM calls in one pipeline?
Yes. You can use the output of one LLM call as input for another. For example, first classify a support ticket, then generate a response based on the classification.
What if the connector produces inaccurate or made-up results?
For critical tasks, reduce the risk of hallucinations by following these best practices:
Configure parameters such as Temperature, Top K, Top P, Frequency Penalty, and Presence Penalty.
Break processes into smaller steps, for example, generate first and verify afterward. This approach provides better control and allows you to validate results before using them.
Create more effective prompts by applying prompt engineering techniques.
Last updated
Was this helpful?