srednoff888-art Agent Browser Automation QA
Choose the smallest browser automation harness that proves the user-facing behavior without hiding failures behind screenshots or model guesses.
Harness Decision
- Use existing project tests first: Playwright, Cypress, Storybook, unit tests, or app-specific smoke scripts.
- Use Playwright MCP or browser DevTools when an agent needs to inspect live pages, DOM, console logs, network requests, or screenshots interactively.
- Use browser-use style controllers only when the task is web navigation across uncontrolled sites and deterministic selectors are unavailable.
- Use screenshot baselines for visual regressions, but pair them with DOM/state assertions for behavior.
- Use manual reasoning only for static pages that do not require runtime interaction.
Validation Loop
- Start the app or open the target URL in an isolated profile when possible.
- Confirm the page is nonblank and the intended route loaded.
- Check console errors, failed network requests, auth redirects, and hydration warnings.
- Exercise the primary workflow with stable selectors or accessible names.
- Capture screenshot evidence only after behavior assertions pass.
- Re-run the check after fixes and record exact commands, URLs, viewport, and remaining risk.
Reliability Checklist
- Prefer semantic locators and accessible names over brittle CSS paths.
- Test desktop and mobile when layout or interaction can differ.
- Keep browser credentials, cookies, and tokens out of logs and screenshots.
- Avoid writing automation that purchases, sends emails, changes billing, deploys, trades, or mutates production data without explicit confirmation.
- Treat AI browser observations as hints until confirmed by DOM, network, trace, or app state.
- Stop and report when captcha, paywall, account lock, robots policy, or legal restriction blocks automation.
Output
Browser QA:
- Harness:
- URL / command:
- Assertions:
- Screenshot / trace:
- Issues found:
- Residual risk: