Ari Weinstein and James Sun, who lead OpenAI’s efforts to build agents capable of operating software, argue that screens and interfaces will remain important as agents improve. They say users will continue to want to see what is happening, make adjustments and retain control, while agents are already becoming more reliable, faster and autonomous than the average human at operating software. These agents combine screenshots, accessibility trees, DOM access and generated code, potentially paving the way for superhuman performance. OpenAI’s Agents API is intended to let these agents work with legacy enterprise systems that lack conventional APIs, with finance and procurement identified as promising initial use cases. The company also presented Dots and plans for asynchronous tool calling at DevDay in September 2026, while interface design and safety controls are expected to evolve alongside agent capabilities.
Source: Read the original article
