Skip to content

Architecture

Overview

┌──────────────┐
│   Webtask    │  Browser lifecycle management
└──────┬───────┘
       │ creates
       ↓
┌──────────────┐
│    Agent     │  Task execution with LLM (text/visual/full mode)
└──────┬───────┘
       │ uses
       ↓
┌──────────────┐
│ TaskRunner   │  Executes steps with tools
└──────┬───────┘
       │ controls
       ↓
┌──────────────┐
│AgentBrowser  │  Page management, context building, coordinate scaling
└──────────────┘

Components

Webtask - Manages browser lifecycle, creates agents

Agent - Main interface with do(), verify(), extract() methods. Supports three modes.

TaskRunner - Executes tasks by calling LLM with available tools

AgentBrowser - Manages pages, builds context (DOM and/or screenshots), scales coordinates

Three Modes

  • text - DOM-based tools (click, fill, type by element ID)
  • visual - Pixel-based tools (click_at, type_text_at by coordinates)
  • full - Both DOM and pixel tools

How it Works

  1. User calls agent.do("task description")
  2. TaskRunner builds context based on mode (DOM, screenshot, or both)
  3. LLM decides which tools to call
  4. TaskRunner executes tools via AgentBrowser
  5. Repeat until task complete or max steps reached

Tools

Common tools (all modes): - goto - Navigate to URL - wait - Wait for time - go_back / go_forward - Browser history - complete_work / abort_work - Task completion

Text mode tools: - click - Click element by ID - fill - Fill form field by ID - type - Type into element by ID

Visual mode tools: - click_at - Click at coordinates - type_text_at - Type at coordinates - hover_at - Hover at coordinates - scroll_at - Scroll at coordinates - drag_and_drop - Drag between coordinates

Stateful Mode

When stateful=True (default), Agent maintains conversation history across do() calls, allowing context to carry over between tasks.