# What is a computer-use agent?

A computer-use agent is an AI agent that operates software the way a person does: it reads a screenshot of the screen and acts through a simulated mouse and keyboard, instead of calling an API built for it. Anthropic introduced this as a public capability in Claude on October 22, 2024, describing it as teaching the model to "use computers the way people do, by looking at a screen, moving a cursor, clicking buttons, and typing text."

The loop is simple even when the task isn't: the model gets a screenshot, decides on one action, a developer's code carries it out, and the next screenshot comes back. That repeats until the task is done or the agent gives up. At launch, Anthropic flagged scrolling, dragging, and zooming as actions the model still handled poorly, since each needs tracking pixel positions across screenshots rather than just reading the page.

OSWorld is the benchmark built to measure how well an agent does this on real multi-app work, not short scripted demos. [Its scores collapse on the longest workflows](/answers/agent-reliability/why-does-success-rate-drop-on-longer-tasks) the same way other long-horizon agent tasks do.

---

Sources:
- Claude 3.5 Sonnet and computer use (Anthropic, October 22, 2024): https://www.anthropic.com/news/3-5-models-and-computer-use (fetched 2026-10-25)

Source: https://tessary.ai/answers/osworld/what-is-a-computer-use-agent
More on Osworld: https://tessary.ai/answers/osworld
From Tessary, agent reliability for AI agents in production: https://tessary.ai