Computer use
Computer use is the class of AI capability where a model operates a computer the way a person does: viewing the screen or interface state, moving a pointer, clicking, typing, and navigating applications to complete tasks. Applied to the web it produces browser agents, currently the most common deployment of the pattern.
How it works
A computer-use loop alternates observation and action: the model receives a representation of the current interface, decides an action, the action executes, and the new state comes back. The observation format is a live design debate with real cost consequences: screenshots are general but token-expensive and lossy about structure; accessibility trees and DOM representations are structured and compact but assume a web target. Recording these loops has the same options, and DOM event streams have the advantage of capturing every intermediate state cheaply enough to keep all of it, which post-hoc debugging and training both want.
Where you encounter it
Model provider capabilities (Anthropic's computer use, and equivalents elsewhere), agent frameworks, and the question of how to record browser-agent sessions.