Browser Tool - Technical Details

Architecture

The Browser tool integration uses: - LocalChromiumBrowser from strands_tools.browser - Playwright for browser automation - PythonAgentTool wrapper to integrate with Strands agent framework

How It Works

Tool Invocation Flow

graph TD
    A[Agent decides to use browser] --> B[Calls browser_wrapper]
    B --> C[Unpacks tool_use object]
    C --> D[Extracts tool_args and toolUseId]
    D --> E[Calls LocalChromiumBrowser.browser with args]
    E --> F[Returns result dict with status/output/content]
    F --> G[Agent processes result]

Wrapper Function

The browser_wrapper in scouts.py handles: 1. Argument injection: Ignores agent and event_loop_cycle_id args from Strands 2. Input unpacking: Extracts input from the tool_use object 3. Result formatting: Wraps results in {status, output, content, toolUseId} format

Action Format

Browser actions are specified as:

{
  "action": {
    "type": "action_name",
    ...parameters
  }
}

Available action types: - init_session: Create a new browser session - navigate: Go to a URL - click: Click an element (requires selector) - get_text: Extract text from element (requires selector) - get_html: Get HTML content - evaluate: Run JavaScript (requires script) - screenshot: Take a screenshot

Integration Points

File: src/influencerpy/core/scouts.py

# Lines 149-194
if "browser" in tools_config:
    browser = LocalChromiumBrowser()

    def browser_wrapper(browser_input, agent=None, event_loop_cycle_id=None, **kwargs):
        # Wrapper logic to handle Strands' argument injection
        ...

    agent_tools.append(PythonAgentTool(
        tool_name=browser.browser.tool_spec['name'],
        tool_spec=browser.browser.tool_spec,
        tool_func=browser_wrapper
    ))

File: src/influencerpy/main.py

Browser tool selection in: - Scout creation flow (line ~666) - Scout update flow (line ~788)

Debugging

Viewing Browser Actions

Check scout logs to see which actions were called:

grep -E "(navigate|click|evaluate)" .influencerpy/logs/scouts/[ScoutName]/[timestamp].log

Common Debug Patterns

Agent uses basic actions only:

tool_use={'action': {'type': 'init_session', ...}}
tool_use={'action': {'type': 'navigate', ...}}
tool_use={'action': {'type': 'get_text', 'selector': 'body'}}

Agent successfully uses evaluate:

tool_use={'action': {'type': 'evaluate', 'script': '...'}}

Agent clicks on element:

tool_use={'action': {'type': 'click', 'selector': '...'}}

Testing

Run verification script:

python3 test_huggingface_papers.py

Expected behavior: - ✅ Navigate to URL - ✅ Extract text content - ❌ Click on links (currently broken) - ❌ Use evaluate for JS (currently broken)