You already know: the research engine (read-only browsing, adversarial pages) and the support agent (actions under policy). This case removes the safety net of read-only: the agent clicks, types and submits. That brings three new problems: perceiving pages reliably, knowing which actions cannot be undone, and never letting page content take control.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud; prices are illustrative.
1. The prompt, and the questions that change the design
A mid-level answer: "the model sees the page and decides what to click". A staff candidate asks: how does the agent see the page (structure or pixels), which actions are irreversible and who approves them, how are logins and payments handled (the agent should never hold passwords or card numbers), and what stops a malicious page from hijacking the task.
Assumptions you state: