Computer use

A capability of agentic AI systems in which the agent perceives the user's screen via screenshots and acts by emitting mouse clicks, keyboard input, and scroll events — operating any application the human can use, without API integration. Originally shipped by Anthropic (Claude with computer-use, October 2024); subsequently adopted by OpenAI (Operator, January 2025) and other vendors.

Computer use is the most powerful and the most dangerous agent capability in 2026 enterprise procurement. Powerful because the agent can operate every legacy application without integration work; dangerous because the action surface is unbounded — the agent can click, type, and submit on any UI, including UIs the deployer didn't authorise. The defensive primitive is action-class containment with screenshot-level audit retention: every click and keystroke is logged with the screenshot that preceded it, and certain action classes (financial, contractual, identity-modifying) require human approval per action. Computer-use deployments that skip the screenshot-level audit substrate cannot meet EU AI Act Article 12 reasoning-trace requirements.

Related frameworks

• MTTD-for-Agents →
• GAUGE →

Articles that analyse this term

• Claude Mythos: what 'too dangerous to release' means for your risk appetite and cyber posture
• Agent incident response: the six-step playbook for when an autonomous-AI deployment breaks production
• Non-human identity for AI agents: the 2026 IAM playbook

Primary sources

• Anthropic. Claude with computer use
• OpenAI. Operator