​
Agent Detection Overview

Superline Agent Detection provides real-time detection of AI agents in browser sessions using metadata and behavioral patterns. This page provides an overview of how the library works and its key components.

​
How It Works

1

Feature Extraction

The library extracts features from browser metadata and user behavior (mouse movements, keyboard patterns, scroll behavior, clicks)
2

Data Processing

Extracted features are processed through standardization, normalization, and one-hot encoding where appropriate
3

Model Scoring

A logistic regression model analyzes the processed features to calculate a detection probability
4

Classification

The probability value is compared to a threshold to classify the session as human or AI agent

​
Key Components

The core component that manages the detection process, including initialization, data collection, and result generation. This is your main point of interaction with the library.
// Example of creating and using the Agent Detector
AgentDetector.init({ debug: false, autoStart: true });
const result = await AgentDetector.finalizeDetection();
Components that collect and analyze different aspects of the browser session. The library includes extractors for browser metadata, mouse events, keyboard events, scroll events, and click events.Each extractor focuses on a specific aspect of user behavior and contributes specific features to the detection model.
A logistic regression model that processes the extracted features to calculate a confidence. The model is trained on a large dataset of labeled browser sessions from both humans and AI agents.The model parameters (weights and intercepts/bias) are included in the library, making it work “out of the box” without requiring server-side processing.
A pluggable system that adapts the library to different operational contexts. The default browser environment provides real-time event collection, while a replay environment is available for offline analysis.
A system for persisting events and metadata during the detection process. The default implementation uses IndexedDB in browser environments to ensure data is available across page loads.

​
Detection Result

When you call AgentDetector.finalizeDetection() or AgentDetector.getCurrentDetectionResult(), you receive a detection result with the following structure: 
{
  isAgent: boolean,     // True if the session is classified as an AI agent
  confidence: number,   // Probability prediction between 0 and 1
  features?: any        // Preprocessed features used for scoring (optional)
}

​
Next Steps

Detection Model

Learn more about how the detection model works

Feature Extraction

Understand how features are extracted and processed