Agent Types

Agentic Browser uses four specialized agents that work together to complete web automation tasks.

Orchestrator

The Orchestrator is the “brain” that coordinates all activity.

Responsibilities

• Task Planning: Breaks your request into actionable steps
• Agent Assignment: Decides which agent handles each step
• Progress Tracking: Monitors completion and handles errors
• Context Management: Maintains state across the task

How It Works

When you submit a task:

1. Orchestrator analyzes your request
2. Creates a plan with ordered steps
3. Assigns each step to the appropriate agent
4. Monitors execution and adjusts as needed
5. Compiles results and responds

Example

Your request:

Find the contact email for Acme Corp and save it to a file

Orchestrator’s plan:

1. Web Surfer → Navigate to acmecorp.com
2. Web Surfer → Find contact page
3. Web Surfer → Extract email address
4. File Surfer → Save to contacts.txt

Web Surfer

The Web Surfer handles all browser interactions.

Capabilities

• Navigation: Go to URLs, click links, use browser controls
• Interaction: Click buttons, fill forms, scroll, select dropdowns
• Reading: Extract text, find elements, capture page state
• Waiting: Handle page loads, AJAX requests, animations

Interaction Types

Element Finding

Web Surfer locates elements using:

• Text content: “Click the ‘Submit’ button”
• Position: “Click the first link in the navigation”
• Attributes: “Fill the email input field”
• Context: “Click ‘Add to Cart’ next to Product A”

Example Actions

Navigate to the login page
→ Web Surfer: goto("https://example.com/login")

Enter username
→ Web Surfer: type("input[name='username']", "john@example.com")

Click submit
→ Web Surfer: click("button[type='submit']")

Coder

The Coder writes and executes code for complex operations.

When Coder Is Used

• Extracting structured data from complex pages
• Transforming or processing collected data
• Handling pagination or infinite scroll
• Custom logic that’s hard to express as browser actions

Capabilities

• JavaScript: Execute in browser context
• Python: Run scripts for data processing
• Data Parsing: JSON, HTML, CSV handling
• API Calls: Programmatic web requests

Example Use Cases

Complex Extraction:

The page has a dynamic table. Coder writes JavaScript to:
1. Find all table rows
2. Extract specific columns
3. Return structured data

Pagination Handling:

Results span 10 pages. Coder creates a loop to:
1. Extract current page data
2. Click "Next"
3. Repeat until no more pages
4. Combine all results

Data Transformation:

Raw data needs processing. Coder:
1. Parses the extracted HTML
2. Cleans and normalizes values
3. Converts to requested format (CSV, JSON)

File Surfer

The File Surfer manages files and documents.

Capabilities

• Download Management: Handle file downloads
• File Reading: Read downloaded content
• File Writing: Save extracted data
• Organization: Manage workspace files

File Operations

Example Actions

Save the PDF
→ File Surfer: download("report.pdf")

Extract text from document
→ File Surfer: read("report.pdf") → extract_text()

Save results
→ File Surfer: write("results.json", data)

Workspace Integration

Files are saved to your Calliope workspace:

• Accessible from other tools (Lab, IDE)
• Persistent across sessions
• Downloadable to your local machine

Agent Collaboration

Sequential Tasks

User: "Download the Q4 report from investor.example.com"

1. Orchestrator → Creates plan
2. Web Surfer → Navigates to site
3. Web Surfer → Finds report link
4. Web Surfer → Clicks download
5. File Surfer → Saves to workspace
6. Orchestrator → Reports completion

Parallel Tasks

For efficiency, agents can work simultaneously:

User: "Get contact info from these 5 company websites"

Orchestrator assigns:
- Web Surfer handles navigation and extraction
- Coder processes and structures data
- File Surfer saves results as they complete

Error Recovery

When something goes wrong:

1. Agent reports the issue
2. Orchestrator evaluates alternatives
3. Different approach attempted
4. User notified if intervention needed

Web Surfer: "Button not found"
Orchestrator: "Try scrolling down first"
Web Surfer: scroll(down) → click(button)

Viewing Agent Activity

Progress Indicators

The interface shows:

• Which agent is currently active
• What action is being performed
• Progress through the plan
• Any issues encountered

Detailed Logs

Click “Show Details” to see:

• Full agent communication
• Exact actions taken
• Element selectors used
• Timing information