Journal / Tools
A convincing screenshot is not a working app
Browser automation gives AI-assisted building an essential feedback loop: did the person’s task actually succeed?
Give the agent a browser
Playwright’s current coding-agent guidance provides browser tooling that agents can use to explore and verify a running application. Its testing documentation emphasizes locators and assertions that wait for the expected state. Together, these tools help move review beyond generated code and into the interface a person actually uses.
Assert outcomes
Choose a task with a clear outcome: complete a form, change a setting, filter a list, or recover from an error. Assert the result that matters. A confirmation heading is stronger evidence than a button having been clicked, and a saved value after reload is stronger than a toast. Use accessible names and roles where possible so the test follows the interface’s meaning rather than its styling.
Our take
Keep exploratory checks and lasting regression tests distinct. Exploration helps discover awkward states; a regression test protects a specific behavior that could break again. Avoid filling the suite with assertions that merely repeat the markup. Include at least one failure path and a narrow viewport for important flows. Screenshots remain valuable for visual review, but they should accompany evidence of behavior, not replace it.
Put it into practice.
Automate your primary user journey and one realistic failure. Review the screenshot, the final state, and any console errors together.
Sources & further reading
Editorial analysis by BEMATE. Sources reviewed 6 September 2026.