05 · Getting Models to Act
Computer Use
Agents that operate a screen
An agent capability where the model operates a computer directly — viewing the screen, moving the cursor, clicking and typing — to use software built for humans, with no special integration required.
Concrete example
Told to file an expense, it opens the web form, fills in the fields, and submits it, the way a person would.
Why it matters
Promises agents that work with any existing software — but it's slow and error-prone today.