🌐 Browser Agent Workflow Builder

Browser Agent Workflow: Weekly Ad Platform Spend Aggregation

Authorization & ToS Check

You are admin on all 4 accounts and authorized to access manually. The ToS-question on agent automation is the genuine grey area. Recommendation:

• Use OFFICIAL APIs where they exist: Google Ads API and Meta Marketing API both expose all the data you need. Free tier covers your volume. Use these instead of browser automation.
• Browser automation only for LinkedIn Ads and Reddit Ads (LinkedIn API is very limited; Reddit Ads API is in beta with gaps).
• For browser-automated platforms, run from your authorized account with reasonable pacing. Do not attempt to bypass any anti-bot measures.

This redirects 50% of your workflow off browser automation entirely, which is the right call. APIs are more reliable than browser agents for structured-data extraction.

Platform Recommendation

Hybrid: Python script for API calls + Anthropic Computer Use for the 2 platforms without good APIs.

Why:

• Computer Use handles MFA, session persistence, and complex SaaS UIs better than headless Playwright in 2026.
• For LinkedIn + Reddit Ads, the data export flows are complex enough that Computer Use's screenshot-based reasoning is more reliable than scraping.
• For Google + Meta, raw API calls are 100× more reliable and don't risk ToS issues.

Manus is overkill for a 1-run-per-week workflow with strict structure. Claude in Chrome would work but lacks the headless scheduling capability you need for an 8am Monday cron.

The Workflow Steps

Phase A: API extraction (parallel, ~2 minutes)

[A1] Google Ads API → fetch last 7 days spend/impr/clicks by campaign
  Verify: response has campaigns, dates align with last 7d, sums are non-zero
  Timeout: 60s
  Failure: log + continue to A2
[A2] Meta Marketing API → same data shape
  Verify: same as A1
  Timeout: 60s
  Failure: log + continue

Phase B: Browser automation (sequential, ~20 minutes)

[B1] Computer Use: open LinkedIn Campaign Manager
  Pre-state: browser open, signed-out
  Action: navigate to campaign-manager URL, sign in (saved cookies if MFA from previous run still valid, else trigger MFA notification)
  Verify: dashboard URL loaded, account-name visible matches expected
  Timeout: 90s for nav, 5min for MFA cycle
  Recovery: if sign-in fails 2x, alert and skip

[B2] Computer Use: navigate to last-7-days view
  Pre-state: dashboard open
  Action: click date-range selector → 'Last 7 days' → apply
  Verify: date range visible at top reads correct range
  Timeout: 30s
  Recovery: if date-range UI moved, take screenshot, re-plan

[B3] Computer Use: export campaign data
  Pre-state: dashboard with last-7-days filter
  Action: click export → choose CSV → download
  Verify: download dialog appears, then file appears in download folder, file size > 0
  Timeout: 60s for the download itself
  Recovery: if export fails, screenshot the page state, alert

[B4] Python: parse downloaded CSV
  Verify: row count > 0, expected columns present
  Failure: alert with the file path

[B5-B8] Repeat B1-B4 for Reddit Ads

Phase C: Aggregate and Post (~3 minutes)

[C1] Python: combine all platforms into unified row format
  Verify: total rows = expected campaign count
[C2] Append to Google Sheet (use Sheets API)
[C3] Format Slack summary message
[C4] Post to #marketing

The 7 Failure Mode Defenses

1. Stale-state: Each Computer Use action verifies state AFTER acting (screenshot + describe). If page state isn't what was expected, re-plan rather than blind-retry.

2. Hidden-element: When clicking 'export' or 'date-range', if the element isn't visible, the agent first scrolls + checks for collapsed menus. Hidden elements re-emerge from accordion menus that were closed.

3. Popup-interrupt: Before each action, the agent screenshots and looks for: cookie banners, 'we've made changes!' modals, MFA prompts, session-timeout overlays. Dismisses each by type. Specifically:

• LinkedIn: known modal 'Try LinkedIn Premium' appears ~30% of sessions. Dismiss with X button.
• Reddit Ads: occasional 'New feature' tour modal. Skip.

4. Viewport-occlusion: Before clicking, scrollIntoView the target element. Use semantic selector (button[aria-label='Export']), not coordinates.

5. Anti-bot-block: Run with reasonable pacing (3-5s between major actions, not 100ms). Use authorized account. If you hit a CAPTCHA, ABORT and alert — don't try to solve. CAPTCHAs are a signal you're being detected as a bot, and bypassing breaks ToS.

6. Navigation-timing: All page loads have explicit waits. Wait for specific elements, not arbitrary timeouts. 'wait until campaign table is visible AND has rows' beats 'wait 10 seconds.'

7. Modal-overlay: Same as popup-interrupt but specifically for ones that block interaction with the page underneath. Always check Z-index; click to dismiss before main flow.

Selector Strategy

Primary: semantic selectors (aria-label, role, text content, data attributes).

Secondary: structural selectors (nth-child, sibling-of-known-element).

Last resort: visual (coordinate-based with relative positioning).

For LinkedIn Campaign Manager:

• Primary: button[aria-label*='Export'], [data-test='date-range-selector']
• Secondary: 'the third button in the toolbar'
• LinkedIn changes its DOM frequently. Plan for selector breakage every 2-3 months.

For Reddit Ads:

• Reddit Ads is more stable; primary semantic selectors usually hold.

Popup Handling Plan

For each platform, maintain a known_popups config:

linkedin_ads:
  - id: try_premium_modal
    selector: '[data-test='premium-upsell']'
    action: click_close_button
  - id: tour_modal
    selector: '.tour-overlay'
    action: click_skip
reddit_ads:
  - id: feature_announcement
    selector: '[role='dialog']:has-text('New')'
    action: click_dismiss

Before each major action, run dismiss_known_popups(). Unknown popups → screenshot + alert + abort that platform's run.

State Checkpoints

After each successful Phase A or Phase B platform:

• Save: platform name, success boolean, output file path or data, timestamp
• Persist to local JSON weekly_ad_run_state_{date}.json
• If subsequent platform fails: workflow can complete with what it has and post 'partial' summary

Resume logic: on workflow restart, check state file. Skip platforms already successful. Resume from first incomplete.

Anti-Bot Considerations

• Run from a residential or datacenter IP that's been used to log in normally before. Don't run from a fresh new IP.
• Use the SAME browser profile/cookies across runs (Computer Use supports persistent profiles). Fresh profiles look more bot-like.
• Pace actions: 3-5s between major actions, longer (10-15s) before/after sensitive operations like export.
• If you see a CAPTCHA: ABORT immediately. Do not try to solve. Alert. Manual remediation required.
• If LinkedIn or Reddit explicitly prohibits agent automation in their ToS (re-check annually), switch to manual collection for that platform.

Implementation Skeleton

/weekly_ad_workflow
  /api_clients
    google_ads_client.py    (uses google-ads-python SDK)
    meta_ads_client.py      (uses Meta Marketing API)
  /browser_agents
    linkedin_runner.py      (Computer Use orchestration)
    reddit_runner.py        (Computer Use orchestration)
    popup_handlers.py       (shared dismiss logic)
  /aggregator
    combine.py              (unified row format)
    sheets_writer.py        (Google Sheets API append)
    slack_poster.py         (Slack webhook)
  /state
    weekly_ad_run_state_*.json (per-run state files)
  main.py                   (orchestrates A1-C4)
  cron.sh                   (called by scheduler at 8am Mon Pacific)

Use a workflow engine (Inngest, simple cron, or AWS EventBridge) to trigger main.py weekly. main.py imports state if exists; runs phases; posts results.

Health Checks & Alerts

• Run start: Slack DM 'Weekly ad run starting'
• Run end: Slack DM with summary (X of 4 platforms succeeded) + post to #marketing if all 4 succeeded
• Run timeout: if workflow exceeds 35 minutes, alert + abort
• No run on Monday by 9:30am Pacific: dead-man switch alerts you (workflow didn't run at all — likely cron issue, not workflow issue)
• Repeated platform failure: if same platform fails 3 weeks in a row, escalate (likely the site changed)

Test Plan

Local validation:

1. Test each API client individually with last week's known data. Verify output matches numbers from manual extract.

2. Test each browser runner against staging-like data (or against real data with ABORT-on-export-success — i.e., go through flow but skip the actual download).

3. Test aggregator with synthetic platform data.

4. Test Slack post in a private channel.

Failure simulations:

• Disconnect network mid-Computer-Use run → verify checkpoint resume
• Trigger a fake popup → verify popup handler dismisses
• Use expired credentials → verify clean abort + alert
• Hit a CAPTCHA (use a VPN to a fresh IP) → verify abort + alert

End-to-end smoke test: run the full workflow on a Sunday afternoon (off-peak). Should complete successfully. Verify Slack post matches manual numbers within 0.5%.

Production rollout:

• Week 1: run in shadow mode (post to private DM, not #marketing). Compare to manual numbers.
• Week 2: if shadow agreement >99%: cut over.
• Week 3+: monitor weekly.

Key Takeaways

• Use APIs for Google Ads + Meta — don't browser-automate when an API exists. This eliminates 50% of your failure surface immediately.
• For LinkedIn + Reddit, Computer Use is the right tool. Manus is overkill for once-weekly; raw Playwright is harder than Computer Use's screenshot reasoning.
• Build popup handling from day one as first-class. All 4 platforms will throw modals at you intermittently.
• Save state checkpoints between platforms. Partial completion (3 of 4 succeed) is acceptable per your spec — but only if the system can post a partial summary cleanly.
• Anti-bot measures are real and increasing. Run from authorized account, residential IP, paced actions. CAPTCHA = abort, not solve.
• This workflow saves 75 minutes/week (90 → 15) at a one-time engineering cost of ~12 hours. Payback in 10 weeks. Worth building, but only if you commit to maintenance when LinkedIn changes its DOM (likely every 2-3 months).