Computer Use in Claude: The Start of Practical AI Agents?

The agent conversation changed from theoretical to practical with https://www.anthropic.com/news/3-5-models-and-computer-use in a way earlier agent framing had not managed from theoretical to practical in a way that earlier agent framing had not managed.

Previous AI agent discussions often described capabilities, planning, tool use, multi-step execution, at a level of abstraction that made them difficult to evaluate concretely. Computer use is different because the capability is specific: Claude can see screenshots, move a cursor, type into fields, and navigate software interfaces. Those are discrete, testable actions that either work or do not on real tasks.

The safeguard question the release raised is the one worth discussing seriously. An AI that can control your browser and desktop applications is operating in your environment with meaningful potential for unintended actions. The gap between well-defined, supervised computer use tasks and fully autonomous operation is the gap that determines whether this is a productivity tool or a risk surface, and that gap is currently large.

Anthropic's framing of computer use as a research capability that requires human oversight rather than an autonomous feature is the honest positioning. The frontier of where that oversight can reasonably be removed is the design question that is still being worked out.

What safeguards should be required before AI agents are allowed to control browsers and desktop apps in a production environment?