Introduction

PYE Agent Curator is a professional desktop application for creating, managing, and executing AI agents based on LangGraph. The tool allows:

Beyond native functionalities, Curator works as an agentic orchestrator (AI RPA), capable of loading and executing externally developed LangGraph code. This allows you to use the Curator interface to monitor logs and manage complex flows created in your own development environment, bringing RPA robustness to Artificial Intelligence.

Getting Started

Download and Installation

The first step to start using PYE Agent Curator is to download the installer for Windows. After downloading, it is essential to install the program by following all the steps indicated by the installation wizard to ensure all dependencies are correctly configured.

2. System Login

When opening the application, you will be presented with the login screen:

To login:

1. Enter your email in the "Email" field
2. Enter your password in the "Password" field
3. (Optional) Check "Stay logged in" to keep the session active
4. Click on "Login"

Action	Description
👁️ Show/Hide password	Click the eye icon to view the password
🔑 Forgot my password	Click the link to receive a recovery email
➕ Create account	Click "Create account" to register

3. License Activation

After logging in, if your account does not yet have an active license, the activation screen will be displayed:

To activate your license:

1. Insert the license key in the format: PYE-XXXX-XXXX-XXXX
2. Click on "Activate Now"
3. Wait for validation (the button will show "Activating...")

4. API Key Configuration

For the agents to work, you need to configure the API keys from the AI providers.

Field	Description	Format
OpenAI API Key	Key for GPT models	Starts with sk-...
Google Gemini Key	Key for Gemini models	Starts with AIza...

3. Main Interface

The main window is divided into functional areas:

4. Dashboard

The Dashboard provides a bird's-eye, real-time view of the health of your agent ecosystem. It consolidates critical metrics so you can proactively monitor efficiency and operational performance.

On the Dashboard, you will find fundamental indicators such as:

5. Creating an Agent

Empty Project

This option allows you to create the basic structure of a LangGraph agent from scratch or integrate Curator with existing code. By selecting a folder that already contains a LangGraph project, Curator acts as an RPA robot, providing the execution interface, log monitoring, and graph visualization for your external code.

1. Click File → New Project (Ctrl + N)
2. Fill in the Name, Path, and Description fields
3. Click "Create Project"

After creation, you will have access to the detailed project settings, where you can define specific execution parameters and behavior for your agent.

Predefined Agents (Document Processors)

Curator includes specialized templates for PDF processing with AI, or custom RPA robot on a solid and scalable structure:

Template	Description
Reframework PDF Batch	Processes multiple PDFs from a folder with a single prompt.
Reframework Excel Batch	Processes multiple Excel files from a folder, the user defines which columns contain information about document paths and prompts that will be used to process the documents
Reframework Base	Base architecture created so the developer can build a custom RPA robot on a solid and scalable structure.

6. Editing an Agent

Editing settings of an existing project:

1. Select the project in the list on the left
2. Click on the menu (...) that appears next to the project name
3. Select "Edit" or "Settings"

For PDF agents, you can modify the prompt, model, validation, and columns.

7. Executing and Monitoring Agents

Start the execution and follow the logs in real-time on the Execution tab.

Button	Description
▶️ Run	Starts agent execution.
⏸️ Pause	Temporarily pauses execution.
⏹️ Stop	Completely stops execution.

8. Graph Visualization

Analyze the structure and flow of your agent dynamically on the Visualization tab.

The main highlight of this tab is the ability to track the flow in real-time. Through fluid animations, Curator highlights exactly which node or step the agent is processing at that moment, allowing unprecedented observability over your LangGraph code execution.

9. Jobs

The Jobs section is the heart of Curator's monitoring. This is where records of all executions performed by your agents are stored. You can track the progress of each task and access the complete interaction history.

Each Job has a status that indicates the operation's result:

Status	Description
Success	Execution was successfully completed.
Cancelled	Execution was interrupted by the user.
Info	Informational status about the process progress.
Failed	A critical error occurred that prevented completion.

10. Triggers

Triggers allow you to automate the triggering of your agents based on external events or time schedules. With them, your agent no longer depends on manual execution and starts operating autonomously.

You can configure different types of triggers, such as Webhooks for integration with other systems, or Schedules (CRON) for tasks that should occur at regular intervals.

In the configuration interface, you define the activation rules, the input parameters that the trigger will send to the agent, and monitor the firing history of each trigger.

11. Logs and Execution History

The Logs tab maintains a complete history of all executions, allowing filtering by project, text search, and export.

Status	Meaning
INFO	Normal processing event
WARNING	Attention, but not critical
ERROR	Something went wrong
SUCCESS	Operation completed successfully

12. ReFramework - Folder

ReFramework Folder is a pre-built agent specialized in batch processing of PDF documents. It was designed to apply a single prompt to all PDFs contained in a specific folder, making it ideal for large-scale standardized processing operations.

How It Works

The agent scans the indicated folder, processes each PDF with the defined prompt, and consolidates all results into a single Excel spreadsheet. The output file is saved in the location specified by the user when configuring the project.

Key Features

Rules and Limitations

Workflow

1. Configuration: Define the input folder, prompt, and Excel output location
2. Execution: The agent processes each PDF sequentially
3. Logging: Each operation is logged in the history
4. Consolidation: Results are saved in the Excel spreadsheet
5. Recovery: In case of interruption, the agent continues from the last processed line

13. ReFramework - Excel

ReFramework Excel is an evolution of ReFramework Folder, offering greater flexibility and granular control over processing. It allows you to define specific prompts for each document or group of documents through a control Excel spreadsheet.

How It Works

The agent reads an Excel spreadsheet where you specify, row by row, which documents should be processed and with which prompts. This allows creating complex and personalized processing flows for different types of documents simultaneously.

Spreadsheet Configuration

To use ReFramework Excel, you need to create a spreadsheet with at least two mandatory columns:

Column	Type	Description
File Path	Required	Full path to PDF file or folder containing PDFs
Prompt	Required	Specific instruction to process the document(s)
Result	Automatic	Filled by agent after processing

Key Features

Operation Modes

1. Single File

Specify the full path to an individual PDF file. The agent will process only this document with the corresponding prompt.

C:\Documents\contract_2025.pdf

2. Complete Folder

Provide the path to a folder. The agent will process all PDFs found in the folder using the same prompt from the row.

C:\Documents\Contracts\2025\

3. Unified Processing

When you specify a folder, the agent reads all PDFs in a unified way, applying the prompt once to the complete set of documents. This is useful for analyses that require context from multiple files.

Rules and Limitations

Workflow

1. Preparation: Create the Excel spreadsheet with Path and Prompt columns
2. Configuration: In Curator, specify which columns to use
3. Execution: The agent processes row by row from the spreadsheet
4. Update: Results are written directly to the spreadsheet
5. Recovery: In case of stop, continues from the row following the last processed

Spreadsheet Example

Path	Prompt	Result
C:\Docs\invoice_jan.pdf	Extract total value and due date	Processed
C:\Docs\Contracts\	List all parties involved	Processed
C:\Docs\report_2025.pdf	Summarize key findings	Pending

14. Agent Validation

One of the most powerful features of PYE Agent Curator's standardized agents is the intelligent validation system. This mechanism ensures the reliability and accuracy of responses generated by AI models, offering two distinct verification modalities.

Validation is especially critical in scenarios where the accuracy of extracted or analyzed data can impact business decisions, regulatory compliance, or downstream automated processes.

1. Exact Double Validation

Exact Double Validation is a deterministic method that compares responses from two different AI models using literal string matching. This method is ideal for structured data extraction where the expected response is objective and does not allow variations.

How It Works

1. Model Selection: User selects two different AI models (e.g., GPT-4 and Gemini 1.5)
2. Parallel Processing: Both models process the same document with the same prompt
3. Response Recording: Responses from both models are saved in separate columns in the spreadsheet
4. Exact Comparison: System compares strings byte-by-byte
5. Validation:
   • ✅ If responses are identical: transaction validated (row remains white)
   • ⚠️ If responses are different: transaction not validated (row highlighted in yellow)

Ideal Use Cases

2. Subjective Double Validation

Subjective Double Validation uses a third-instance AI approach to evaluate semantic consistency between two responses. This method is suitable for analyses involving interpretation, summarization, or extraction of complex concepts.

How It Works

1. Primary Model Selection: Selection of two models to generate initial responses
2. Parallel Processing: Both process the same document/prompt
3. Validator Invocation: A third AI (validator) receives both responses
4. Semantic Analysis: The validator evaluates:
   • Information consistency
   • Factual discrepancies
   • Logical contradictions
   • Significant divergences
5. Decision:
   • ✅ If validator considers responses consistent: transaction validated (white row)
   • ⚠️ If validator finds discrepancies: transaction invalidated (yellow highlighted row)

Ideal Use Cases

Method Comparison

Aspect	EXACT Validation	SUBJECTIVE Validation
API Calls	2 (both models)	3 (models + validator)
Speed	⚡ Faster	🐢 Slower
Cost	💰 Lower	💰💰 Higher
Response Type	Objective and structured	Interpretive and contextual
Criterion	Literal equality	Semantic consistency
Best For	Structured data	Complex analyses

Interpreting Results

When a transaction is marked with a yellow row, it indicates that validation detected inconsistencies. In this case, it is recommended to:

1. Manual Review: Analyze both responses from primary models
2. Check Source Document: Consult the original PDF to confirm the correct information
3. Adjust Prompt: If there are many discrepancies, the prompt may need to be more specific
4. Consider PDF Quality: Scanned or poorly formatted PDFs can generate inconsistent results

15. Support and Help

PYE Agent Curator is a software in constant development. For this reason, your feedback and reporting of any unexpected behavior are fundamental for us to implement new innovations and continuous improvements, as well as rapid fixes.

For technical questions or reporting issues, please contact: