The agent's eyes are /screenshot, its hands are /eval

The agent is supposed to navigate a vendor portal that has no API. You give it two URLs: GET /screenshot?fullPage=true and POST /eval.

The agent calls /screenshot, gets a base64 PNG, sends it to a vision model, asks "where is the export button." The model answers with coordinates or a selector. The agent calls /eval with document.querySelector(...).click(). Then /screenshot again to confirm the modal opened.

The whole loop is HTTP. No Playwright SDK loaded into the agent process. No browser driver to keep alive. The model doesn't know what a CDP session is. Two endpoints and a vision call, repeated until the export downloads.