1. The agent wires deterministic API/auth fixtures into your app behind a local env flag (the README includes a setup prompt for this), so runs don't depend on live services. 2. The agent declares flows in a small JSON manifest — clicks, fills, text checks, and layout assertions like "no horizontal overflow" or "this button is in the viewport" — and runs them via the CLI at desktop/tablet/mobile viewports. 3. Each run captures screenshots and writes a local HTML report the agent can read to verify and fix its own work. Exit code 0/1, so the same command works in CI.

It's a thin layer on Playwright: the goal isn't to replace it, but to make the see-act-verify loop simple and deterministic enough that an agent can set it up and iterate on it unattended. No account, no hosted service, no Storybook. Example output in the README.