Use them as fast drafters inside a workflow that keeps a human approval step. Give the model real context (a working seed test, the requirement or spec, your conventions), ask for a plan before code, then run and read every generated test. Playwright's official test agents split the job into a planner that writes a Markdown test plan, a generator that turns it into test files, and a healer that repairs failures; practitioner guides advise reviewing each stage, opening a pull request for healer changes rather than letting it commit to main, and never treating generated output as final until it has run. Check that assertions test behavior and not just that the page loaded, and keep sensitive data out of prompts.

Sources: Playwright Test Agents, QASkills guide, PractiTest 2026 State of Testing