Desktop

Desktop is the space that touches the real machine. It is built as perception first and action second, which is the opposite of how most automation is built and the reason it is meant to degrade more gracefully.

Perception before action

Before anything is clicked, the screen is captured and analysed — text extracted, elements identified. The plan is formed against what is actually on screen rather than against coordinates recorded when someone wrote the script. A layout change degrades the match instead of silently clicking the wrong thing.

Plan, execute, validate

Desktop work is orchestrated as a handoff: a planning stage decides the steps, an execution stage performs them, and a validation stage confirms the screen ended up where it was supposed to. Write actions require confirmation and an external receipt in the registry, which is what makes a failed automation reportable rather than merely finished.

What the Brain can reach

The Brain has two desktop capabilities: one hands a desktop skill to a coordinating agent, the other drives a visible browser through an agent in OpenFang. Both target running agents. Neither has been verified end to end.