Building Computer use for Cloud Agents

AI coding agents can now visually verify their work through a feature called computer use, allowing them to see a screen, click, type, and scroll. This is achieved with a two-level agent architecture where a primary agent delegates GUI tasks to a specialized subagent that operates in a tight loop of analyzing screenshots and executing actions. A model-agnostic protocol translates high-level actions from different LLMs into a set of generic atomic actions, ensuring client-side simplicity and cross-provider compatibility. The entire system runs securely on the Oz platform in isolated cloud sandboxes with virtual displays, enabling agents to be triggered remotely without accessing a user's personal machine.