Skip to content

Browser Tool (Experimental)

⚠️ Status: The Browser tool integration is currently EXPERIMENTAL.

Overview

The Browser tool uses Playwright-based browser automation via strands_tools to enable scouts to navigate web pages and extract content dynamically. This is more powerful than simple HTTP requests for JavaScript-heavy sites.

Current Limitations

🔴 Multi-Step Workflows Not Reliable

The AI agent currently struggles with complex, multi-step browser interactions:

  • ❌ Click Actions: The agent often doesn't click on links even when instructed
  • ❌ Form Filling: Multi-field forms are not reliably completed
  • ❌ Navigation Sequences: "Navigate → Find Element → Click → Read Content" workflows often fail
  • ❌ JavaScript Evaluation: The agent rarely uses the evaluate action even when prompted

Why?: The underlying model (gemini-2.5-flash by default) tends to take the shortest path and only uses basic actions like navigate and get_text.

✅ What Works

  • ✅ Single Page Navigation: Navigate to a URL and read its content
  • ✅ Text Extraction: Extract visible text from a page
  • ✅ Simple Content Scraping: Get HTML from static pages
  • ✅ Basic Screenshots: Capture page screenshots

✅ Good Use Cases

  • Extracting content from a known, stable URL
  • Reading blog posts or article pages
  • Monitoring specific web pages for text changes
  • Simple content aggregation from listings
  • Complex multi-page workflows
  • Dynamic content requiring clicking through UI
  • Sites requiring form interactions
  • Workflows needing JavaScript inspection

Example: What Works

# ✅ This works well
scout = {
    "name": "TechCrunchDaily",
    "type": "web",
    "url": "https://techcrunch.com",
    "tools": ["browser"],
    "goal": "Extract the titles and summaries of the top 3 articles"
}

Example: What Doesn't Work Reliably

# ❌ This often fails
scout = {
    "name": "HuggingFaceTop",
    "type": "web",
    "url": "https://huggingface.co/papers",
    "tools": ["browser"],
    "goal": "Find paper with most upvotes, click it, and read the abstract"
}
# Problem: Agent won't click on the paper link

Workarounds

If you need complex browser automation:

  1. Use Direct URLs: If you know the exact URL of the content, navigate directly to it
  2. Pre-scraping: Use a separate script to find links, then feed URLs to scouts
  3. Simplify Goals: Break complex workflows into multiple simpler scouts
  4. Alternative Tools: Use RSS, Reddit, or Arxiv scouts when applicable

Known Issues

Issue Status Workaround
Agent ignores click instructions Open Navigate directly to target URL
Agent ignores evaluate for JS inspection Open Use simpler selectors
Agent doesn't follow multi-step prompts Open Use single-action goals

Future Improvements

We're actively working on: - [ ] Better prompt engineering for multi-step workflows - [ ] Testing with more capable models (GPT-4, Claude) - [ ] Custom high-level browser actions (e.g., click_most_upvoted_paper()) - [ ] Example-based few-shot prompting

Reporting Issues

If you encounter issues with the Browser tool:

  1. Check the scout logs: .influencerpy/logs/scouts/[ScoutName]/
  2. Look for which browser actions were actually called
  3. Note if evaluate or click actions are missing
  4. Report to: GitHub Issues

Contributing

Help us improve the Browser tool! We especially need: - Test cases for working multi-step workflows - Prompt templates that successfully trigger clicking - Model comparisons (which models follow instructions better?)