An agent can write a program in minutes and still be unable to try it: the moment the program asks for a key press or draws a screen, the agent is outside, looking at a closed door.
Tools made for hands and eyes
Interactive tools - pickers, menus, editors, whole graphical programs - are made for a person at a keyboard. An agent that built one cannot check it works; an agent that needs one cannot use it. So a person is called in to press keys and read the screen, mid-task, by hand.
Two doors, and a driver
Every tool keeps two doors: the interactive one for people, and a structured one - plain arguments, output a program can read - for agents and scripts. Most of the time an agent walks through the second.\n\nFor checking the first door itself, a terminal driver starts the program, presses keys, waits until the screen shows what it should, and reads the screen as text. Interactions written down as scenes - press this, expect that - run on every change, and their screens become evidence.
A ramp beside the stairsA building with a ramp beside its stairs is not a different building. Everyone can enter; each by the way that suits them.
Called, not copied
Like the package conveyor, it is called: a project declares its scenes and calls the conveyor at a pinned version. The same driver is offered to agents over the Model Context Protocol, the open standard through which programs such as Blender and Photoshop already open themselves to agents.
What it does not do yet
Graphical programs come later, through Linux's accessibility interface or each program's own connector. A scene proves what it was written to expect, and no more.
Where it stands
Researched and planned; nothing built. It follows two heavier pieces of work and then begins with the driver, tried first on ocoeco's own picker - the tool whose real-terminal check could not be made.