The installer asks a question; the agent reads the screen

A vendor's CLI ships a setup wizard. Arrow keys, a license prompt, "press Enter to continue." No flags, no unattended mode. Humans have been clicking through it since 2014.

Your onboarding agent opens a terminal session and drives the wizard through the terminal kit's automation endpoints. GET /api/v1/terminal/snapshot returns the rendered screen as text. GET /find locates "Accept license" with a regex. POST /press sends Enter. POST /wait blocks until the next prompt renders. The agent reads, decides, presses, and repeats, against the same libvterm state a human would see in the browser tab.

Nobody wrote an expect script. Nobody patched the vendor. The wizard stays a wizard, and the agent is just another user at the keyboard, one who reads the whole screen at once.