A powerful, emerging use case for computer use agents is to have them test the software they've just written. Instead of a human performing QA, the agent can interact with the application's UI, verify functionality, and catch design issues, effectively closing the development loop from code generation to validation.
Unlike previous browser-in-the-cloud agents, OpenAI's "DOTS" personal assistants are each provisioned with their own persistent Linux virtual computer. This architectural choice is significant because it allows the agent to run full desktop applications, not just web browsers, greatly expanding the scope of tasks it can perform.
OpenAI trains its models on its own proprietary "computer use harness." By providing this harness directly within the Agents API, they offer developers a potential speed, cost, and accuracy advantage. The model is already optimized for this specific environment, making the official API more performant than a custom-built harness for many use cases.
To improve performance for long-running personal agents, OpenAI is moving beyond basic caching. They are developing features like guaranteed 12-hour cache windows and "pre-warming," where developers can pay to populate the cache with an expected prompt ahead of time. This treats agent state more like a pre-computable asset.
The goal for computer use agents has shifted beyond mimicking human actions to exceeding them in speed. The primary bottleneck is no longer the AI's reasoning but the real-world latency of the software it operates, like website loading times. This changes how developers must think about agent performance optimization.
OpenAI's Decisions API wasn't a new model. It was rapidly prototyped by engineers using existing Luna model weights, focusing on inference optimizations like parallel batching for structured outputs. This demonstrates how frontier labs can quickly replicate competitor features by leveraging their existing model stack and a "hacker culture."
OpenAI's "App Shots" feature provides AI models with a deep understanding of an application's interface by leveraging accessibility APIs. This technology, originally designed for screen readers for visually impaired users, gives the AI structured data about UI elements that a simple screenshot would miss, making it a powerful input modality.
A key evolution in computer use agents is their ability to move beyond single, sequential actions. Agents now write and execute chunks of JavaScript code to perform multiple operations at once. This significantly improves speed and capability, allowing them to handle more sophisticated, long-horizon tasks on websites and applications.