Computer use agent

Computer use agents operate graphical interfaces instead of only calling structured APIs. They typically receive screenshots, decide on actions such as click, drag, type, or keyboard shortcut, execute those actions in a sandboxed environment, and repeat until the task is complete. They are powerful for legacy apps and visual workflows, but they create unique risks because webpages, images, and UI text can contain instructions that conflict with the user's goal. Safer deployments use containers or virtual machines, minimal privileges, domain allowlists, sensitive-data restrictions, and human confirmation for consequential actions.