Where AI really shines

Spot the test-authoring tasks where AI is the obvious tool — pattern-following, typing-heavy work where the agent does the labour and you do the deciding.

You watched the agent extract a POM and a fixture. Both followed the same shape: clear template, concrete input, mechanical output. That's the sweet spot — and once you spot the shape, it's everywhere.

More places it shines

Same shape, different jobs:

  • Split a long spec file. A 600-line file with twenty test(...) blocks across four features needs grouping. The agent infers the natural seams (by feature, by user role, by happy/error path) and proposes a split.
  • Migrate to semantic locators. Sweep tests/ for page.locator('#…') / page.locator('.…') calls, open the matching page in playwright-cli, grab the getByRole / getByLabel equivalent, and rewrite.
  • Parameterized variants. One happy-path test plus a list of states ("logged out", "no items", "expired token"); the agent emits eight variants from the same shell.
  • Test data factories. Hand the agent observed API responses from a session and it produces typed factories with sensible defaults.
  • Accessibility assertions. Sweep the snapshot for missing labels, headings, alt text — fold into expect(...).toHaveAccessibleName(...) checks.
  • Test descriptions. Read a dense flow, propose a clearer test name. Boring, fast, useful.
  • Coverage filling. "Read tests/checkout/. List scenarios from the spec doc that aren't covered." Then pick what's worth adding yourself.

Where it doesn't shine

These are still your job:

  • What to assert. The agent picks something — usually the cart counter. Whether that is the assertion that catches the regression you care about is a judgment call.
  • What to abstract. A POM with one caller is overhead. A fixture that hides login changes the meaning of every test that uses it. Abstractions are taste calls; the agent has none.
  • What to test at all. The hardest decision in testing isn't writing the test — it's knowing which behaviour deserves one. The agent can't smell what matters in your product.

Agents draft. You ship.

You made it 🎭

And that's the workshop! The AI block sits on top of the basics and writing-tests groups, and you've now seen both ends. You should now be able to:

  • Record, read, and refactor a Playwright test
  • Use UI mode and the trace viewer to debug anything
  • Pick locators that survive a redesign and assertions that don't race
  • Hand a browser to your coding agent through CLI
  • Generate tests from a prompt
  • Lift repeated locators into POMs and repeated setup into fixtures with the agent doing the typing
  • Recognize the test-authoring jobs where AI pays back tokens — and the ones where it doesn't

Where to go next:

Thanks for spending the day with me. Go test something. 🎭

Hands on

Spot more places to send the agent

Exercise 1 of 1

Enjoy the rest of the day!

And we'll see each other tomorrow at the conference! 🎉