Playwright Best Practices For Agents

Playwright best practices

Condensed, opinionated guidance for writing Playwright tests that are readable, isolated, and resilient — built for coding agents, around Playwright's agent CLI (playwright-cli) and its no-GUI debugging flows. Maintained by Checkly — the same practices apply whether you run these tests in CI or as production monitors.

Load a reference file from references/ only when the task needs it (see routing table). Each reference ends with links to the full /learn articles for depth.

Scope: all guidance assumes the @playwright/test test runner with TypeScript — its test, fixtures, projects, config, and web-first expect. Examples are TypeScript (.spec.ts); the same APIs work in JavaScript. It does not target the standalone playwright automation library (which has no test runner, fixtures, or auto-retrying assertions). Imports are import { test, expect } from '@playwright/test'.

The agent CLI is what makes this skill shine. Playwright's agent CLI — playwright-cli, package @playwright/cli — is a separate, token-efficient, no-GUI browser you drive command by command to discover locators and step through failing tests. It's distinct from the standard npx playwright CLI, and the Agentic workflow below leans on it throughout. → references/debugging.md

Core rules (always apply)

1. Locator priority: prefer user-facing locators — getByRole > getByLabel / getByPlaceholder / getByText > getByTestId > CSS/XPath. CSS/XPath tie tests to implementation and break easily. → references/locators.md
2. Web-first assertions: use auto-retrying expect(locator).toBeVisible() / toHaveText() etc. Never assert on a one-shot value you pulled out manually (innerText() then toBe). → references/assertions.md
3. No hard waits: never waitForTimeout(). Trust auto-waiting actions and web-first assertions; for explicit waits use waitForURL / waitForLoadState / waitForResponse. Avoid networkidle. → references/waiting.md
4. Isolated & independent: each test sets up its own state and can run in any order, in parallel. No test depends on another. Provision state via API in setup, not through the UI. → references/test-structure.md, references/flakiness.md
5. One feature per test: if a test's assertions span more than one feature, split it. Keep tests short and focused.
6. Reuse auth, don't re-login: sign in once, persist storageState, reuse it across tests via a setup project. → references/auth.md

Routing table

Agentic workflow (no GUI)

The interactive tools — --ui, --debug (Inspector), show-trace — are GUIs you can't drive. Author and debug through the non-interactive signals instead.

Having playwright-cli available is highly encouraged — both phases below lean on it. Confirm with playwright-cli --version and install it if missing — npm install -D @playwright/cli, then run it via npx playwright-cli (or install globally with npm install -g @playwright/cli to call playwright-cli directly). Everything still works without it, but you lose the inspect/verify loop and fall back to guessing.

Author — discover, don't guess. Read locators off the live page rather than from source: playwright-cli open <url> → playwright-cli snapshot prints the accessibility tree — the roles and accessible names that power getByRole/getByLabel — so you author the user-facing locator straight from what it shows. → references/locators.md

Run & debug:

1. Run and read stdout: npx playwright test path/to/file.spec.ts. The reporter prints the failing assertion and the call log — which locator/assertion timed out and what Playwright actually saw. Read it; don't guess.
2. Read error-context.md: on an expect failure Playwright writes an aria-snapshot of the page at the moment it failed to the test's test-results/.../error-context.md. This is machine-readable page state — open it to see what was actually rendered. (Playwright ≥ 1.60)
3. Capture artifacts, not GUIs: add --trace on to drop trace.zip into test-results/ for inspection.
4. Step through it live with playwright-cli (no GUI): run npx playwright test path/to/file.spec.ts --debug=cli in the background — it pauses and prints a session name. Then playwright-cli attach <session-name> and drive it: playwright-cli snapshot (page state + element refs), playwright-cli step-over, playwright-cli console error, playwright-cli network, playwright-cli eval "…". Inspect why the locator didn't resolve or what actually rendered, then fix and re-run. (needs the agent CLI; full detail in references/debugging.md)
5. Fix the root cause (usually a locator, a missing web-first assertion, or a hard wait), then re-run until green. Don't paper over flakiness with retries — see references/flakiness.md.

Full agentic-debugging detail (the playwright-cli discovery and --debug=cli stepping workflow) is in references/debugging.md.

Stay current. These primitives are recent and version-gated — check npx playwright --version and playwright-cli --version, and update both packages if they're behind. Detail in references/debugging.md.