The Visual Builder lets you create agents using a drag-and-drop interface. Instead of working step-by-step in text, you can place and connect nodes on a canvas to define how your workflow runs.

Each node represents a task — such as document parsing, RAG question answering, data extraction, or running a model. You can easily configure each step using simple forms, set inputs, and control what outputs are passed to the next step.

The Visual Builder makes it easy to see how everything is connected and how data moves through your workflow. You can rearrange steps, zoom in and out, and quickly adjust connections as needed.

Once your workflow is ready, you can run it directly, export it, or switch to the step-based editor for further edits.

The Visual Builder excels at creating complex workflows with branching logic, conditional execution, and parallel processing paths, as these structures are more easily understood and modified in graphical form. The interface includes features for zooming, panning, rearranging nodes, and validating connections to ensure data flows correctly between steps. Once the visual workflow is complete, it can be executed directly from the builder, exported as a JSON configuration file, or further refined using the step-based editor.

3.1 Builder overview

The canvas represents the full execution flow of an agent, from input to final output. Each block is a node, and connections define how data moves between steps.

All components are fully visual and configurable directly in the builder.

3.2 Left panel (node types)

The left sidebar contains the core building blocks that can be dragged onto the canvas:

• Input Defines how data enters the agent (for example text, image, or file input).

• Node General-purpose processing steps that pass data forward.

• Classifier Routes execution based on intent or classification logic.

• Bot Allows user to include a Bot interface in the agent.

• Condition Adds branching logic based on rules or outputs.

• Transformer Allows user to specify which data to use in the agent process as it moves to the next step.

3.3 Agent Configurations:

• Files: This section is used to upload, manage, and associate data assets with the agent.

• Services: It serves as a service catalog for the agent. It determines which capabilities are available when building workflows.

• Settings: Allows to set Agent global configurations from selecting models to overall control in the agent.

3.4 Canvas controls

On the canvas, the following actions can be performed:

• Nodes can be dragged to reposition them
• Nodes can be connected to define execution flow
• Nodes can be selected to edit their instructions and configuration

• Note: The number of the node indicates the order in which the User added that node on the canvas, and does not indicate the order in which the services or the node will be activated. The order of agent execution follows the order of the nodes and how they are linked to each other, rather than the node number.

3.5 Zoom and utility actions

The bottom-left controls allow the following operations:

• Zoom in
• Zoom out
• Clear the canvas

3.6 Action buttons

Below the utility buttons, quick access to key actions is provided:

• Save Agent Saves the agent definition as a JSON file. This file can later be uploaded to instantly recreate the agent.

• Run Executes the agent with the current configuration.

• Home Returns to the main dashboard.

4. Input Node

The Input Node defines how data enters the agent workflow. It is typically the starting point on the canvas and determines the format and structure of the data that downstream nodes will receive.

Each Input Node can be configured by selecting an input type and adding a short description to guide users on what to provide. This makes the agent easier to use and ensures consistent input formatting.

Available Input Types

Notes

• The selected input type directly impacts how downstream nodes interpret and process data.
• Clear descriptions improve usability, especially when agents are shared or reused.

5. Node

The Node represents a general-purpose processing step in the agent workflow. It is used to execute a specific service and pass the result forward to the next connected node.

Each Node acts as a bridge between data and capability, allowing the agent to perform operations such as reasoning, transformation, retrieval, or external service interaction.

Service Selection

Every Node includes a “Choose service” dropdown. This list is dynamically populated from the Services section in the agent configuration.

• Only services that are added and enabled in the Services section will appear here
• This ensures that the Node operates within the capabilities explicitly defined for the agent
• Services can include model-based operations, data processing tools, integrations, or custom logic

Instruction Field

Each Node provides an Instruction field where you define how the selected service should behave.

• Instructions guide the execution of the service
• They can include prompts, rules, or task-specific directions
• Clear and precise instructions improve output quality and consistency

Custom Datasets as Services

In addition to predefined services, custom datasets can also be used within a Node:

• Datasets must first be configured in the main Datasets section [ROHAN: Add link here to Datasets]
• Once created, datasets must then be added as a file in the agent process in the Files section
• Once listed, the dataset and each of the columns specified in the Dataset creation become selectable in the transformer node
• To use dataset-related Agent workflow services, select All-Datasets in the Services catalog (left side nav in visual editor)
• This allows workflows to directly interact with structured or domain-specific data with dataset-related services

Key Characteristics

• Nodes process incoming data and produce outputs for downstream steps
• They can be chained to create multi-step workflows
• The behavior of a Node is fully determined by the selected service and its instruction
• Nodes are reusable and can be reconfigured without affecting the overall structure

Notes

• If a required service is not visible, ensure it has been added in the Services section
• Keep instructions concise but explicit to avoid ambiguity
• Nodes can be combined with conditions and transformers to build complex logic flows

6. Classifier Node

The Classifier Node is used to categorize or label input data based on predefined classification types. It enables routing, filtering, and decision-making within workflows by assigning structured outputs such as labels, categories, or extracted information.

Classifier Selection

Each Classifier Node includes a “Choose classifier” dropdown with a fixed set of available classifiers.

• These classifiers are not dynamically generated from the Services section
• However, corresponding capabilities may still be configured or supported through Services if needed
• The dropdown provides a consistent and standardized set of classification options
• Select the Classifier you wish to choose in the Service Catalog (left side nav in visual editor) and click ">"
• This enables the Classifier node to update when you select and place it on the canvas workspace

Available Classifiers

Key Characteristics

• Outputs are structured and can be used for conditional routing
• Helps in building branching logic when combined with Condition nodes
• Works well with both raw and preprocessed inputs
• Can be chained with other nodes for multi-step analysis

Notes

• Since classifiers are predefined, customization is limited to how their outputs are used downstream
• For advanced behavior, combine classifiers with Nodes and Transformers
• Ensure the selected classifier aligns with the expected output format for the next step in the workflow

7. Bot Node

The Bot Node is used to integrate and execute a pre-configured bot within the agent workflow. It allows you to delegate specific tasks to a reusable bot that has already been defined with its own logic, instructions, and capabilities.

Bot Selection

Each Bot Node references a bot selected from the Embedded Agent Bots available in the Settings section.

• The list includes both pre-built bots and user-created bots
• User-created bots are those configured in the Bots section
• Only bots that are properly set up and available in settings can be selected

This ensures that all bot executions are consistent with their predefined configurations.

How It Works

• The Bot Node receives input from previous nodes - user is encouraged to link a transformer with the appropriate data state (i.e. Agent State)
• It passes the input to the selected bot
• The bot processes the request based on its internal configuration
• The output is returned and passed to the next node in the workflow
• The user is then able to query the bot with the knowledge it has received in the workflow
• If using the bot in the middle of a workflow, the user must interact with the bot and indicate that they are finished interacting with the bot before the agent workflow will continue to the next steps

Key Characteristics

• Encapsulates complex logic into reusable components
• Reduces duplication by reusing existing bot configurations
• Maintains consistency across workflows using the same bot
• Supports both simple and advanced multi-step reasoning within a single node

When to Use

• When a task has already been defined as a reusable bot
• When you want to standardize behavior across multiple workflows
• When delegating complex reasoning or interactions to a dedicated component

Notes

• Changes made to a bot in the Bots section will reflect across all Bot Nodes using it
• Ensure the selected bot is properly configured before using it in a workflow
• Bot Nodes can be combined with Classifiers and Conditions for dynamic execution paths

8. Condition Node

The Condition Node introduces decision-making into the workflow by evaluating a condition and routing execution based on the result. It enables branching logic, allowing the agent to follow different paths depending on the data it receives.

Value Selection

The “Choose value” dropdown determines what data the condition will evaluate.

Default Values

• agent-state Represents the current internal state of the agent
• user-document Refers to the input provided by the user, especially in document-based workflows
• none Used when no predefined value is required

Conditional Operators

The second dropdown defines how the selected value is evaluated. The available operators are:

Define Condition

The Define condition field is where the comparison value or expression is specified.

• Can be a static value, keyword, or expression
• Works in combination with the selected operator
• Should align with the data type of the selected value

Outputs

The Condition Node has two possible execution paths:

• True (Green output) → Followed when the condition is satisfied
• False (Red output) → Followed when the condition is not satisfied

Key Characteristics

• Enables branching and control flow within the agent
• Works with both predefined and dynamic values
• Integrates seamlessly with outputs from previous nodes
• Supports simple comparisons as well as advanced evaluations

Notes

• Ensure the selected value exists in the workflow before using it in a condition
• Use clear and predictable outputs from previous nodes to avoid ambiguity
• Combine with Classifier Nodes for more intelligent routing decisions

9. Transformer Node

The Transformer Node is used to access, extract, and reshape data from different stages of the agent’s execution. It enables you to work with intermediate outputs and reuse them in downstream steps.

Unlike standard processing nodes, Transformers focus on state access and data transformation, making them essential for building flexible, multi-step workflows.

Purpose

Transformers allow you to:

• Retrieve data from any step in the agent’s process
• Reuse outputs without recomputing them
• Restructure or prepare data for the next node

For example, if an agent workflow has 5 steps, a Transformer can pull data from any intermediate state (step 1, 2, 3, etc.) and pass it forward for further processing.

Input Selection

The “Choose input” dropdown defines which data source the Transformer will use.

Default Inputs

• agent-state Provides access to the internal state of the agent across all steps
• user-document Refers to the original input provided by the user
• none Used when no predefined input is required

Dynamic Key Inputs

Dynamic Key inputs may include:

• rag_answer
• rag_sources
• agent_report
• description
• Any other fields produced during execution

These dynamic options are created based on previous agent activity and allow Transformers to integrate seamlessly with the evolving data flow of the agent.

How It Works

1. The Transformer displays a specific input (state or output)
2. It extracts or reshapes the data
3. The transformed result is passed to the next connected node depending on user selection and intent

Key Characteristics

• Enables access to intermediate and final outputs
• Supports data reuse across multiple steps
• Decouples data retrieval from processing logic
• Improves modularity and flexibility of workflows

When to Use

• When you need to reference outputs from earlier steps
• When preparing data for another node (e.g., Bot, Classifier, Condition)
• When working with multi-step or stateful agent flows
• Extremely useful in dealing with CSVs or other data structures when filtering, sorting, or searching prior to continuing to a next agent state/node of activity

Notes

• Ensure the selected input exists in the workflow before using it
• Use meaningful outputs in earlier nodes to simplify transformations
• Transformers are especially powerful when combined with Conditions and Nodes for dynamic execution paths

10. Files

The Files section is used to upload, manage, and associate assets with the agent.

Capabilities include:

• Uploading documents, images, tables, datasets, or zipped sources
• Assigning file types such as document, image, dataset, table, or source
• Reusing existing datasets or sources already available in the workspace

Uploaded files become available as contexts that can be selected by agent services such as parse_document, build_table, or vision. Multiple files can be added before saving and exiting.

11. Services

The Services section defines all the capabilities available to an agent. These services power the execution of Node, Classifier, Bot, and Transformer components in the workflow.

Only services that are added and enabled in this section will be available for selection inside nodes. This ensures controlled, predictable, and modular agent behavior.

Core services

General building blocks for common tasks such as chat, retrieval, extraction, and logic control.

Examples:

• chat – conversational responses
• rag_answer – retrieval-augmented answers
• vision – image understanding
• ocr – text extraction from images
• web_search – online search retrieval
• extract – structured field extraction
• answer – direct question answering
• prompt_builder – dynamic prompt creation
• agent_report – report generation
• embedded_bot – embedded execution
• condition – branching logic
• boolean – rule evaluation

Classifiers

Lightweight text analysis tools for labeling or scoring content.

• sentiment – positive/negative tone detection
• emotions – emotional classification
• topics – topic categorization
• tags – keyword tagging
• intent – intent recognition
• ratings – scoring or grading
• ner – named entity recognition
• summary – concise text summaries
• category – predefined grouping
• q_gen – question generation

Datasets

Tools for preparing, querying, and analyzing structured data.

• select_keys - selects specified keys from a JSON dictionary
• build_dataset – create datasets from JSON
• ds_command_filter - applies filter commands to a dataset
• ds_column_filter - keep rows where a selected column meets your condition
• ds_quick_stats – generates a statistical report based on selected column
• ds_column_analysis - generates a report based on selected column
• ds_report - generate a report of the dataset and the workflow results based on the agent run
• ds_column_select - returns the selcted column from the dataset
• ds_ask_dataset - use a natural language question to retrieve relevant information from the dataset
• ds_readout - returns the text from a set of rows from the dataset for display
• ds_smart_filter – find rows that match the meaning of your query
• ds_keyword_filter - filter rows based on exact text matches in the selected column
• dataset_plot – visualize data
• ds_statistics – perform deeper statistical analysis and generate insights
• load_dataset – load saved datasets
• create_json – provide a list of agent keys to consolidate into a new JSON dictionary
• ds_stat_analysis - generate statistical analysis of input data csv file

Specialized services

Advanced utilities for targeted or complex workflows.

• build_table - create table from CSV data
• query_custom_table - database look-up in natural language
• json_extractor - converts a text chunk with embedded json into a structured dataset element
• semantic_filter – meaning-based filtering
• text_filter – rule-based text filtering
• document_filter – document-level filtering
• table_filter – structured table filtering
• load_kb - load knowledge base into agent state used in 'ask_kb' calls
• ds_ask_kb - answers knowledge base questions from a dataset input
• transformer – text transformation tasks
• aggregate_context - provide a list of context names to consolidate
• parse_document – convert documents to text files
• create_context – build reusable context blocks from the most relevant passages in a source based on query
• report_commentary – generate commentary of key process results from the agent-state - no input context required
• speech_gen – generate audio output from text
• image_gen – generate images from text
• get_stock_summary - stock ticker look-up (requires internet access)
• speech - transcribe a speech file
• speech_batch - transcribe a collection of speec
• vision_batch - answer question based on a collection of image files
• parse_batch - create source from document batch
• extract_tiny - extracts a key-value pair
• website_scraper – extract web content from allowed websites (note: many websites prevent this)
• extract_table – extract tables from documents based on query

Integrations

Connect the agent to external systems or hosted models.

Examples include storage services, email, and external model providers.

Custom services

Workspace-specific or user-defined services added for specialized use cases.

Below is the list of supported services, their expected instruction formats, descriptions, and applicable context sources.

Conclusion

The Visual Builder establishes a clear and structured approach to agent design by translating complex workflows into an intuitive visual format. By organizing execution into interconnected nodes—covering inputs, processing, decision-making, and transformation—it enables users to construct both simple and highly sophisticated agents within a single, unified canvas.

Throughout the system, emphasis is placed on modularity, reusability, and controlled capability exposure. Components such as Nodes, Classifiers, Bots, Conditions, and Transformers work together to create flexible pipelines, while the Services and Files sections ensure that only explicitly configured resources and capabilities are used. This not only improves maintainability but also promotes consistency across different agent implementations.

Key mechanisms like Condition nodes introduce dynamic branching, and Transformer nodes provide access to intermediate execution states, allowing workflows to evolve beyond linear processing into adaptive, state-aware systems. At the same time, reusable bots and services reduce duplication and streamline development.

Finally, the use of concise instruction fields and UI-oriented labels maintains clarity for end users without disrupting the overall documentation style. Together, these elements position the Visual Builder as a scalable and user-friendly framework for designing, managing, and executing agent workflows with both precision and flexibility.