AI Web Task Runner

What does AI Web Task Runner do?

AI Web Task Runner is an Apify Actor that turns natural-language browser tasks into controlled Playwright automation runs.

It can:

• browse public websites
• follow task-relevant links
• extract structured results
• capture screenshots
• save raw HTML
• generate a human-readable report
• export a reusable Playwright Python script from the successful task trajectory

This Actor is designed for public-web automation, extraction, research, and script generation.

It is not a login bot, spam bot, comment bot, messaging bot, or anti-bot bypass tool.

How it differs from a fixed scraper

Most scrapers are built for one website and one output shape.

AI Web Task Runner is different:

• you describe the task in natural language
• the Actor opens one or more start URLs
• it follows task-relevant public pages
• it records an action trajectory
• it extracts best-effort results even without an LLM
• it can optionally use an LLM for safer planning, schema mapping, and summarization

This makes it useful for a wider class of public-web tasks than a single-purpose scraper, while still staying controlled and safety-constrained.

Main modes

run_task

Default mode.

Use this for general browser-task execution, such as:

• finding features
• locating pricing information
• summarizing a product page
• finding the correct public page for a business task

extract

Use this for structured extraction.

If you provide an extractionSchema, the Actor tries to map observed content into that schema.

research

Use this to browse task-relevant public pages and produce a summary with source URLs.

generate_script

Use this to run a task and export a reusable standalone Playwright Python script based on the successful action trajectory.

audit_lead

Optional compatibility mode.

This preserves the previous lead/contact-audit workflow and outputs company-profile style results for website contact and outreach auditing.

Input examples

Example 1: Pricing extraction

{
  "task": "Find the pricing plans and extract plan name, price, billing period, and main features.",
  "startUrls": [
    { "url": "https://example.com" }
  ],
  "mode": "extract",
  "extractionSchema": {
    "plans": [
      {
        "name": "",
        "price": "",
        "billingPeriod": "",
        "features": []
      }
    ]
  },
  "maxPages": 5,
  "captureScreenshots": true
}

Example 2: Research task

{
  "task": "Find what services this company offers and summarize them with source URLs.",
  "startUrls": [
    { "url": "https://example.com" }
  ],
  "mode": "research",
  "maxPages": 6,
  "maxDepth": 2,
  "sameDomainOnly": true
}

Example 3: Generate reusable Playwright script

{
  "task": "Open the website, navigate to the pricing page, and extract the pricing table.",
  "startUrls": [
    { "url": "https://example.com" }
  ],
  "mode": "generate_script",
  "generateReusableScript": true,
  "maxPages": 5,
  "captureScreenshots": true
}

Example 4: Optional lead audit template

{
  "task": "Audit this website for contact and sales outreach readiness.",
  "startUrls": [
    { "url": "https://example.com" }
  ],
  "mode": "audit_lead",
  "maxPages": 5,
  "captureScreenshots": true
}

Output record types

For all main modes except audit_lead, the default dataset contains only:

task_result

One final task-level record. This is the main export row and the recommended unit for pricing and CSV export.

Detailed page snapshots and extracted items are stored in the key-value store and referenced from the final task result.

It contains:

• task
• mode
• final status
• pages visited
• steps executed
• summary
• result payload
• confidence
• screenshot keys
• report key
• trajectory key
• generated script key if applicable
• page snapshots key
• extracted items key

audit_lead compatibility output

When mode = audit_lead, the default dataset contains:

• company_profile

The older page records are preserved as compatibility artifacts in the key-value store, not as billable default dataset rows.

Key-value store artifacts

For the main task-runner modes, the Actor saves:

• REPORT.html
• TASK_RESULT.json
• TASK_TRAJECTORY.json
• PAGE_SNAPSHOTS.json
• EXTRACTED_ITEMS.json
• GENERATED_SCRIPT_RECORD.json if enabled
• generated_script.py if enabled
• generated_script_metadata.json if enabled
• screenshots if enabled
• raw HTML if saveHtml = true

This means the default dataset stays clean and export-friendly, while detailed execution artifacts remain available in the key-value store.

For audit_lead, compatibility artifacts include:

• COMPANY_PROFILES.json
• PAGE_RECORDS.json
• REPORT.html
• run_log.json

Generated Playwright script

When generateReusableScript = true or mode = generate_script, the Actor saves:

• generated_script.py
• generated_script_metadata.json

The generated script:

• is standalone Playwright Python
• is based on the recorded safe action trajectory
• includes comments for the reproduced browser steps
• contains no secrets
• does not rely on arbitrary LLM-generated executable code

Safety model

The Actor only allows a fixed safe action set:

• visit_url
• click_link_text
• click_css_selector
• type_text
• press_key
• select_option
• wait
• extract_current_page
• collect_links
• stop

The Actor does not allow:

• arbitrary Python execution
• shell execution
• unrestricted JavaScript execution
• login/cookie automation in V1
• posting, commenting, or messaging automation
• paywall bypass
• CAPTCHA bypass
• destructive actions
• purchases or form submissions that change state

Deterministic vs LLM-assisted mode

Deterministic mode

If llmProvider = none, the Actor still works.

It will:

• open start URLs
• collect page snapshots
• infer task keywords
• follow obvious task-relevant links within limits
• extract visible text, headings, links, tables, prices, emails, phones, and structured candidates
• produce a best-effort task result

LLM-assisted mode

If an LLM provider is configured, the Actor may use the model for:

• safe action planning
• choosing the next allowed action
• mapping page snapshots into extractionSchema
• summarizing findings

The LLM is constrained to structured JSON outputs and validated before use.

If the LLM fails, the Actor falls back to deterministic behavior.

Optional lead-audit template

The older lead/contact-auditor behavior is still available in:

"mode": "audit_lead"

That mode keeps:

• lead-audit heuristics
• page-level contact findings
• company profile aggregation
• compatibility outputs for outreach workflows

It is now an optional template mode, not the main identity of the Actor.

Example use cases

• Extract pricing plans from a SaaS website
• Extract a table from a public webpage
• Research product features from a company website
• Find contact or sales paths from a public website
• Summarize same-domain public pages related to a topic
• Generate a reusable Playwright script for a repeated browser task

Limitations

• This Actor is intended for public websites.
• It does not support login-heavy or state-changing automation in V1.
• Deterministic extraction is heuristic and best-effort.
• Some websites hide key data behind scripts, forms, or client-side UI patterns.
• LLM mode can improve planning and summarization, but it is still constrained and fallback-safe.
• It is not an anti-bot bypass product.

Troubleshooting

I only got the homepage

Increase maxDepth or start from a more relevant public page.

The final result is incomplete

Increase:

• maxPages
• maxDepth
• timeoutSeconds

and consider using extract mode with an extractionSchema.

The Actor did not visit the page I expected

Check:

• TASK_TRAJECTORY.json
• PAGE_SNAPSHOTS.json
• EXTRACTED_ITEMS.json
• REPORT.html

These artifacts show what the Actor actually saw and did.

I want the older contact-audit behavior

Use:

"mode": "audit_lead"

The Actor should not submit forms or log in

That is expected in V1. The safety model deliberately avoids state-changing automation.