OpenAI is building AI agents that can operate software on their own. Two of the people leading that work say the screens we click on today will stick around anyway.
Ari Weinstein and James Sun made that argument on a recent episode of The a16z Show, released on October 26, 2026. Their reasoning is simple: people like interacting with things directly, and that preference does not vanish because a machine can do the clicking for them.
It is a notable position from a company whose agents, by the speakers’ own account, can already outpace the average human at running software. Think of it as a self-driving car company insisting the steering wheel stays.
What the agents can do now
The conversation centered on computer-using agents, or CUAs. These are AI systems that operate software the way a person does: reading the screen, moving through menus, filling in fields, and finishing tasks.
According to the podcast discussion, these agents have improved on three fronts: reliability, speed, and autonomy. As of October 26, 2026, they had reached a point where they can operate software more reliably and quickly than average human users.
The speakers credited a blend of techniques for the progress. Agents now draw on screenshots, accessibility trees, DOM access, and generated code to get work done.
Those terms deserve unpacking. A screenshot lets the agent see the screen like a person would. An accessibility tree is a structured map of the on-screen elements, originally built so assistive tools could describe an interface. DOM access gives the agent a view into how a web page is put together underneath. Generated code lets the agent write its own small programs when that is faster than clicking around.
The discussion suggested this combination could push agents toward superhuman performance in the near future. That is framed as a possibility, not an accomplished fact.
Who is making the case
Weinstein is the product and engineering lead for OpenAI’s Computer Use effort. Before joining OpenAI, he co-founded Sky, built by the company known as Software.inc.
Sun handles enterprise browser and agent development at OpenAI. He also came to the company after founding a startup of his own.
The podcast built on announcements from OpenAI’s DevDay in September 2026. That event featured Dots, an Agents API, and plans for async tool calling.
The Agents API matters most for this story. It lets computer-using agents work with legacy systems that lack traditional APIs.
For the non-developers: an API is a clean, official doorway that lets one program talk to another. Plenty of older business software never got one. An agent that can simply use the screen like a human sidesteps that problem entirely.
Why the interface survives
Weinstein and Sun argued that user preference for direct interaction will keep interfaces relevant. People want to see what is happening, adjust it, and feel in control.
The speakers described UI design as something that will keep changing as agents develop, with human preferences and agent capabilities existing side by side.
The conversation also touched on safety and control measures as agents gain more ability.
What this means for software and the people who build it
The most immediate stakes sit in the enterprise. The research tied to the podcast pointed to finance and procurement as areas where agents could work through complex, traditional software environments.
OpenAI is pursuing both sides of that equation. It has consumer-facing products like Dots alongside enterprise applications, suggesting the company sees agents and interfaces developing together rather than one replacing the other.
The things to watch are concrete. One is whether async tool calling moves from plans to a shipped feature. Another is how quickly enterprises plug the Agents API into legacy systems that have resisted automation for years.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.