Computer-Use Agent
Also called computer use, GUI agent, browser agent, browser use.
A computer-use agent operates a graphical computer interface the way a person does, by looking at screenshots and issuing mouse clicks, keystrokes, and scrolls, so it can work with software that has no API.
Description
The agent loop alternates between a screenshot and an action until the task is done. Benchmarks such as OSWorld measure performance on real desktop tasks. Because these agents read arbitrary on-screen content, they are exposed to prompt injection from web pages and documents, so they are run in isolated environments with human confirmation for sensitive actions.
Sources
- Xie et al. (2024). OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments.
- Anthropic (2024). Computer use tool.
Cite this entry
Protologue. (2026). Computer-Use Agent. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0144). https://protologue.com/t/computer-use-agent/
BibTeX
@misc{protologue_computer_use_agent,
title = {Computer-Use Agent},
author = {{Protologue}},
year = {2026},
howpublished = {Protologue: A Taxonomy of Prompting and LLM Techniques, v1.4.0},
note = {Entry PTL-0144},
url = {https://protologue.com/t/computer-use-agent/}
}New terms, as they're coined
One email a week with the newest techniques, each defined and sourced like this entry.
Delivered through Beehiiv. Unsubscribe from any issue.